MCP tool reference
Every tool the Clinical Terminal MCP server exposes: health-tech vendors, the hospitals and health systems that buy from them, the people who run both, and the signals moving between them, plus the few tools that publish a page. To connect a client first, start with the setup guide.
Each entry lists the tool’s parameters, an example arguments object, the fields that come back and the errors it can return. Access comes with a Landscape subscription, which covers every tool here except the two Analysis tools; those are staff-only.
Start here
Two tools answer most first questions. Reach for these before chaining anything.
For a question about one company, get_vendor_dossier takes a name and returns the profile, the hospital customers with the verbatim quote each was evidenced by, current products and recent signals — in one call, instead of four.
For a question about a market, find_vendors_for_problem takes a buyer problem in plain words and routes it to one of the Clinical Terminal market segments, then returns that segment’s vendors ranked by traction.
One habit worth forming: an empty result here means not covered, never does not exist. Several tools say so explicitly in their response rather than handing back a bare empty list.
How responses work
- Every tool returns one text content block holding JSON. Most responses are pretty-printed; get_vendor_dossier, get_provider and get_persona return compact JSON.
- A missing key means unknown. Null and empty fields are dropped from every row rather than sent as null, so never read an absent field as zero or false.
- A response longer than 50,000 characters is cut off and ends with … [truncated at 50000 chars — narrow your query with filters or a smaller limit]. The two dossier tools keep to their own smaller budgets instead and list what they trimmed in meta.truncated.
- Arguments are validated against each tool’s schema before it runs. A wrong type, an out-of-range number or an unknown enum value is rejected with a validation error, so the tool-specific errors below are the ones a well-formed call can still hit.
- Numeric limits accept any number in range and are rounded down to a whole number.
Free job search (no key)
Current jobs at health-tech vendors and their investors, searchable from Claude Code without an account or a key.
A separate, free endpoint answers three questions: what is open, what a posting says, and who the employer is. Jobs come straight from each employer’s own job board (Greenhouse, Lever, Ashby, Workday and others), checked daily, and only postings seen in the last 14 days are listed. Each answer links to the employer’s Clinical Terminal page.
Add it to Claude Code with one command:
claude mcp add --transport http clinicalterminal-jobs https://mcp-jobs-687925922613.us-central1.run.app/jobs/mcp- search_jobs — keyword search with filters for remote, seniority, department, employer and product category. At most three postings per employer are listed unless you name the employer.
- get_job — one posting in full: description, pay where the employer publishes it, and the apply link.
- get_employer — a short brief on a hiring vendor: what it does, its best-evidenced customers, leaders and recent news.
Ask it: “Find remote product manager roles at ambient documentation companies.”
“What does Abridge do, and who are its customers?”
“Show me senior engineering jobs in Boston with a posted salary.”
Limits are per network address: 30 calls a minute and 500 a day, of which 300 may be employer briefs. Claude Code only: Claude Desktop and claude.ai custom connectors call from shared servers, so their users would share one set of limits. Posting text comes from employers’ job boards, with email addresses and phone numbers removed.
Companies & vendors
The published health-tech vendors, their products, and their own websites. Coverage is uneven: nearly every vendor has a logo and a summary, fewer have a product catalog, so a thin result means thin data rather than a thin company.
get_vendor_dossier
The fastest path to any single-vendor question. One call by name returns the profile, board, hospital customers with their evidence quotes, current products, recent signals, ownership and what we know of the vendor’s website — instead of chaining a search, a fetch, a product lookup and a signal query.
Ask it: “Give me the dossier on Waystar.”
| Name | Type | Description |
|---|---|---|
| vendor required | string | A vendor name (“Waystar”), vendor id or numeric slug. A unique near-miss spelling is corrected automatically. 1–200 characters. |
{
"vendor": "Waystar"
}One object. Each section can independently be replaced by an unavailable marker (see errors).
| vendor | The vendor profile, as get_vendor returns it. |
| board | The confirmed members of the governing board in role order (chair first, then CEO directors), each with role and person id, plus boardSize (voting members), the sources that confirm them (the SEC Form D filing and its date, or the vendor’s own board page), unmatchedOnFiling and notShown for directors we can’t name, and mayBeOutOfDate after 12 months. null means no board on record yet, not that the company has none. |
| customers[] | Up to 10 customer relationships, strongest evidence first, with the evidence quote, evidence tier and citable article. |
| products[] | Up to 5 current products. |
| signals[] | Up to 5 recent signals, one per source article. |
| ownership[] | Acquisitions in either direction. Omitted when none are on file. |
| funding | Funding rounds and funding news. When we hold rounds that cite a source (SEC Form D filings and the funding announcements we cover), roundsTracked is true and rounds[] lists each with its amount, date, investors and source links; an amount marked sold is SEC dollars sold, which can differ from the announced figure. Otherwise roundsTracked is false. events[] always carries the funding and M&A news we hold. |
| siteMap | Pages we know on the vendor’s website, by category, with freshness. |
| meta | The query, the resolved vendor, spellingCorrected when a typo was fixed, and truncated listing anything trimmed to fit the budget. |
| upcomingEventAppearances | Up to five upcoming conferences the subject appears at, with roles. Absent when none are recorded; from news and LinkedIn, so absence is not evidence. |
- Ambiguous name — Returns candidates[] (up to 5, each with id, name and domain) instead of a dossier. Call again with the id you meant.
- Unknown vendor — Entity not found: "<id>". followed by how to pass an id: the id from search_vendors, a numeric slug, or a terminalId. A vendor name is not an id.
- A section failed — That section becomes { unavailable: true, reason }; the rest of the dossier still comes back. Retry, or use the section’s own tool.
search_vendors
Find vendors by name, domain or category. Multi-word queries match on all words, then top up with looser any-word matches; results rank exact name matches first, then by how much news a company is generating.
Ask it: “Which vendors work in denials management?”
| Name | Type | Description |
|---|---|---|
| query required | string | Matched against name, domain and categories; multi-word phrases work. 1–200 characters. |
| limit optional | integer | Maximum results, 1–50. Default 20. |
{
"query": "denials management",
"limit": 10
}An array of vendors, best match first.
| id | Vendor id — pass it to get_vendor. |
| name | Vendor name. |
| domain | The vendor’s web domain. |
| summary | Company summary. |
| logoUrl, logoSource | A logo you can use now. stored is our permanent copy; scraped is the source link, which may later expire, so copy it rather than linking to it. Absent when no logo is stored yet. |
| companyType | Always vendor. |
- No match — Returns { results: [], note } instead of an array. The note is the coverage contract: a miss means Clinical Terminal does not cover the vendor, not that it does not exist.
get_vendor
The full profile once you have an id or numeric slug, including boardSize and chairName (null when no board or chair is on record). There is no vendor score on this server — the denial/appeals/prebill scores were retired from every read path on 2026-08-04 and are never returned.
| Name | Type | Description |
|---|---|---|
| id required | string | Vendor id from search_vendors, or the numeric slug. A vendor name is not an id. |
{
"id": "100000001"
}One vendor object.
| id, name, domain | Identity. |
| slug | Numeric terminalId, as a string. |
| summary, logoUrl | Company summary and logo URL. |
| categories[] | Segment slugs, rolled up from the products this vendor sells. |
| foundedYear, headquarters | Company basics, when known. |
| ceoName, ceoTitle | The current CEO, when known. |
- Unknown or unpublished id — Vendor not found.
find_vendors_for_problem
Capability discovery across healthcare through the Clinical Terminal market ontology (17 domains, 163 segments, 1,071 capabilities) — the same router the app uses. Routing is deterministic, never a model guess. Vendors whose own websites show the matched capability come first, then vendors in the same segment; each one carries a quote.
Ask it: “Which vendors do ambient clinical documentation?”
| Name | Type | Description |
|---|---|---|
| problem required | string | A buyer problem or capability in plain words, or an exact segment name from list_segments. At least 2 characters. |
| limit optional | integer | Maximum vendors, 1–25. Default 10. |
{
"problem": "denials management",
"limit": 10
}An object with status: "ok" and the segment it routed to.
| segmentId, segmentName, domainId, domainName | The segment the problem routed to, and its domain. |
| capabilities[] | Up to three matched capabilities, best first. |
| matchMethod | How it routed: perspective, name, typo or search. |
| correction | Set when a misspelling was corrected — pass it on to the user. |
| vendors[] | Ranked vendors: id, name, terminalId, domain, summary, logoUrl, matchedOn (capability or segment), matchedCapabilities, capabilityEvidence, segmentEvidence and up to two evidence quotes with their source URL. |
| coverageNote | What an empty or short list means. Pass it on. |
- Could not route the phrase — Returns status: "no_route", sometimes with didYouMean[]. Call list_segments, pick the closest segment and call again with its exact name.
- Reading the evidence counts — They count labelled website features, not customers. Use get_relationships for customers.
list_segments
Every segment of the market ontology with its domain and the number of published vendors labelled into it — the map behind the tool above. A zero count means no labelled vendor yet, not an empty market.
None — call it with an empty arguments object.
{}An array of all 163 segments, including empty ones.
| id | Segment slug. |
| name | Display name — pass it to find_vendors_for_problem. |
| domainId, domainName, description | The segment’s domain and what it covers. |
| vendorCount | Published vendors whose websites are labelled into the segment. |
search_products
Search what vendors actually sell. Defaults to current products, because a large share of the catalog is discontinued; status always comes back so you can tell the difference.
| Name | Type | Description |
|---|---|---|
| query optional | string | Matched against product name and description. 1–200 characters. |
| vendor_id optional | string | Restrict to one vendor’s catalog. |
| include_inactive optional | boolean | Also return discontinued products. Default false. |
| limit optional | integer | Maximum results, 1–50. Default 20. |
{
"query": "prior authorization",
"limit": 20
}An array of products, by vendor name then product name.
| id | Product id — pass it to get_product. |
| vendorId, vendorName | The vendor that sells it. |
| name | Product name. |
| productType | service, software, platform, module or device. |
| status | active or inactive. |
| description | Product description. |
| segmentIds[] | Market-ontology segment slugs the product is tagged with (the ids list_segments returns). |
- Empty array — Usually means the vendor is not catalogued, not that it sells nothing.
get_product
One product in full, with the vendor it belongs to.
| Name | Type | Description |
|---|---|---|
| id required | string | Product id from search_products. |
{
"id": "prod_8f2k1"
}One product object, with the same fields as a search_products row.
- Unknown id — Product not found.
get_vendor_pages
A vendor’s own website as data — the pages we have crawled and classified. Filter by category to jump straight to the page worth reading.
| Name | Type | Description |
|---|---|---|
| vendor required | string | Vendor name, id or numeric slug. 1–200 characters. |
| category optional | enum | One of homepage, about, leadership, contact, news, careers, blog, legal, product, pricing, customers, integrations, demo, resources, other. Omit for every page. |
| include_dead optional | boolean | Also return pages that no longer resolve. Default false. |
| limit optional | integer | Rows to return, 1–100. Default 50. |
| offset optional | integer | Rows to skip, for paging. Default 0. |
{
"vendor": "Waystar",
"category": "customers",
"limit": 25
}One object.
| vendor | The resolved vendor: id, name, terminalId. |
| pages[] | Each page: page_id (pass it to get_page_content), url, classification, scraped (whether we hold its text), and isDead when it no longer resolves. |
| pagesKnown, pagesReturned | Live pages we hold, and rows in this response. |
| scrapedCount, deadPages | Pages we hold text for, and pages that no longer resolve. |
| lastCrawledAt | When content was last fetched — the freshness stamp. |
| note | Why a result is empty or partial, and how to page on. |
- Ambiguous name — Returns candidates[] instead of pages.
- Unknown vendor — Entity not found: "<id>". followed by how to pass an id: the id from search_vendors, a numeric slug, or a terminalId. A vendor name is not an id.
- Unpublished vendor — Its website corpus is not available on this server.
- No crawl yet — An empty list with a note saying we have not mapped the site — absence of data on our side, not absence of pages on theirs.
- Corpus tools switched off — An error saying the website-corpus tools are temporarily disabled; every other tool is unaffected.
get_page_content
Read the text of one of those pages: product copy, case studies and bios rather than a guess from the URL. It arrives inside an explicit untrusted-content fence, because it is verbatim third-party marketing — evidence to quote, never instructions to follow.
| Name | Type | Description |
|---|---|---|
| page_id required | string | A page_id from get_vendor_pages. A URL or a vendor name is not a page_id. |
{
"page_id": "pg_3k9x2"
}One object.
| page_id, url, classification | Which page this is. |
| vendor | The published vendor that owns the page. |
| scraped | Whether we hold its text. |
| title, content | When scraped: the title, and Markdown up to 10,000 characters inside a BEGIN/END UNTRUSTED WEBSITE CONTENT fence. |
| truncated | True when the content was cut at a paragraph boundary. |
| scrapedAt | When the text was fetched. |
- Not scraped — Not an error: scraped: false with a note to fetch the URL directly.
- Unknown page_id — An error naming the id and where page_ids come from.
- Page of an unpublished vendor — Its content is not available here.
get_help
The subscriber guide (what Clinical Terminal is, where each product question goes, how to set up your agent, set your territory on the Map and get alerts) plus your own setup status. Your assistant can explain the product and tell you exactly what is not set up yet. It cannot change settings; it sends you to the page that does.
Ask it: “How does Clinical Terminal work, and is my account set up?”
None — call it with an empty arguments object.
{}Two text blocks: the guide as Markdown, then one JSON object.
| guideVersion | A short hash of the guide text, so you can tell which version is live. |
| connection | Which account and key is answering: account (the email), and key with itsname, prefix, tier and the day it was created. An unnamed key carries a fix pointing to Settings, where you can name it. |
| account.territory | set (with the states), nationwide, not_set or unknown. |
| account.agentSetup | confirmed, in_progress, not_started or unknown. |
| account.newsAlerts | on, off, not_set_up or unknown, with how many organizations you follow. |
| fix | On any gap: the app page to open and what to click. |
- A part of the account could not be read — That part reads unknown, which does not mean the setting is off; the rest still answers.
get_my_context
The organizations you have been working with in Clinical Terminal chat, most recent first. This is what lets an assistant resolve “that vendor” or “my target” across sessions instead of asking you again.
Ask it: “What was I looking at last time?”
None — call it with an empty arguments object.
{}One object.
| recentEntities[] | Up to 25, most recent first: tid (usable with get_relationships or get_vendor_dossier), name, type, lastSeenAt and seenCount. |
| updatedAt | When the chat memory last changed. |
- No chat history — Plain text, not JSON: no chat memory yet, or the key has no linked user.
get_my_brief
Your morning brief for a day, the same one your Clinical Terminal homepage shows: the themed summary, the LinkedIn posts behind it, upcoming events, and the people it mentions. Read-only; it never generates a brief.
Ask it: “What's in my morning brief today?”
| Name | Type | Description |
|---|---|---|
| date optional | string | An Eastern-time day, YYYY-MM-DD. Defaults to today in America/New_York. A brief covers the 24 hours before it was built (the nightly run is 1 am ET). |
{
"date": "2026-10-24"
}One object.
| status | ok, or error when the AI summary failed (the posts are still the day’s real posts), or not-generated when no brief exists for that date. |
| brief / summary / themes | The written brief, a one-line summary, and the themes (announcements, hiring, events, product, other). Machine-written from public posts. |
| posts[] / upcomingEvents[] | The posts the brief was built from, and events in the next 30 days with attendees. |
| people[] | People with a photo on file: name and terminalId (usable with get_person). |
| mentionedNames[] | People the brief mentions who have no photo on file, by name only: no id, no link. |
- No brief for that date — status: not-generated with the reason: not built yet today, a past day with no brief, or a future date. That is not the same as a quiet news day.
- Key has no linked user — status: no-user.
get_daily_brief
Generates the Daily Brief Clinical Terminal chat writes: a written brief on one day’s healthcare-vendor LinkedIn posts, shaped to your persona. Optionally framed for a conversation with one person, or narrowed to a topic. Each call is a fresh AI run, limited per key per day; for the brief already made this morning, use get_my_brief.
Ask it: “Write me a brief on yesterday's vendor news, focused on prior auth.”
| Name | Type | Description |
|---|---|---|
| date optional | string | An Eastern-time day, YYYY-MM-DD. Defaults to today. The brief draws on that day’s 100 highest-priority posts. |
| person_id optional | number | A person’s terminalId (from search_people or get_person). Frames the brief as preparation for talking to them. |
| focus_topic optional | string | Narrows the brief to a topic, e.g. “prior authorization”. 1–120 characters. |
{
"date": "2026-10-24",
"focus_topic": "prior authorization"
}One object.
| status | ok; no-posts (no LinkedIn posts that day); error (the write-up failed but the posts are real); budget (daily limit reached). |
| brief | The written brief. Machine-written from public posts. |
| subject | The person the brief was framed for, when you passed person_id. Name only if they have no photo on file. |
| posts[] / upcomingEvents[] | The posts it drew on, and upcoming events. |
| people[] / mentionedNames[] | As in get_my_brief: people with a photo as name and terminalId; people without one by name only. |
- Daily limit reached — status: budget with the limit. It resets at midnight UTC.
- Unknown person or bad argument — status: not_found or invalid_argument with a message saying which.
Providers
The buy side: US hospitals, health systems, FQHCs, rural health clinics, ASCs, imaging centers, physician groups, skilled nursing and home health.
get_provider_segments
Start any provider-market question here. Every segment with its entity count, its trailing-7-day signal volume, and the signed change against the week before. All segment keys come back, including the quiet ones, so you never have to guess whether a segment exists.
Ask it: “Which parts of the provider market moved this week?”
None — call it with an empty arguments object.
{}An array with one row per provider segment, zero-filled.
| providerType | Segment key — pass it to search_providers. |
| count | Provider entities in the segment. |
| signals7d, signals7dPrior | Signal rows in the last 7 days, and days 8–14. |
| delta | The signed change between the two. |
- Reading signal counts — They count article-by-entity rows, not distinct articles, and are non-zero only for hospital and health-system segments.
search_providers
The provider counterpart to vendor search — by name, segment, state, health system or size. Names collide heavily in this data, so filter by state or size when you can.
| Name | Type | Description |
|---|---|---|
| query optional | string | Provider name or fragment; multi-word phrases work. 1–200 characters. |
| provider_type optional | enum | A segment key from get_provider_segments, such as health_system, acute_care, critical_access, fqhc or asc. |
| state optional | string | A two-letter code or a full state name (“TX” or “Texas”). |
| health_sys_id optional | string | Only members of one health system — a healthSysId from a result. |
| min_beds optional | integer | Minimum staffed beds, 1–5000. Drops providers with no bed count, such as clinics. |
| sort optional | enum | size (most beds first), name, or signal_recency (newest news first). Defaults to relevance with a query, name without. |
| limit optional | integer | Maximum results, 1–50. Default 20. |
{
"query": "Baylor Scott",
"state": "TX",
"sort": "size",
"limit": 10
}An array of providers.
| id, terminalId | Either one works with get_provider. |
| name, city, state | Identity and location. |
| providerSubtype, companyType | Segment key, and the coarse type it derives from. |
| staffedBeds | Bed count, for hospitals and health systems. |
| healthSysId | The parent health system, for a system-to-members lookup. |
| LinkedIn URL. | |
| logoUrl, logoSource | The provider’s logo, as on search_vendors: stored or scraped. Absent means no logo is stored yet, not that the provider has none. |
| lastSignalAt | Most recent news detection, for hospitals and health systems. |
- Unrecognised state — Unknown state: "<state>". with the list of valid codes.
search_service_lines
Hospitals and health systems that run a clinical department — cardiology, oncology, neurology, orthopedics and 41 more — each with the evidence that says so: staff job titles naming the department, and sections of the provider’s own website. Use it instead of search_providers for any department question; that tool matches names only. This is a floor, not an inventory: a provider missing from the list may still run the department.
Ask it: “Which hospitals in Texas run a cardiology program?”
| Name | Type | Description |
|---|---|---|
| service_line required | enum | The department key, such as cardiology, oncology, neurology, orthopedics, obgyn, behavioral, imaging or emergency. |
| state optional | string | A two-letter code or a full state name (“TX” or “Texas”). |
| cities optional | string[] | Only providers in these cities; needs state. A metro is several cities, so list the suburbs too, such as ["Houston", "Sugar Land", "The Woodlands"]. Case, “St” versus “Saint” and punctuation don’t matter. |
| level optional | enum | hospital (individual facilities), system (health-system parents) or any. A system’s beds are its system-wide total. Default hospital. |
| provider_type optional | enum | One segment, such as acute_care, critical_access or childrens. |
| min_beds optional | integer | Minimum staffed beds, 1–5000. Drops providers with no bed count. |
| min_evidence optional | enum | strong keeps only a dedicated website section or 2+ staff titles; any also includes single mentions. Default any. |
| include_unclassified optional | boolean | Return providers with no recorded segment, shown with a null providerType. Most are real hospitals and regional systems; a few are universities or payers, so check the name. false withholds and counts them. Default true. |
| limit optional | integer | Maximum providers, 1–50. Default 25. |
{
"service_line": "oncology",
"state": "TX",
"cities": [
"Houston",
"Sugar Land",
"The Woodlands",
"Katy"
],
"limit": 25
}One object: the line, the providers largest first, and counters that explain every provider not returned.
| line | The department key and its label. |
| results[].terminalId, openInGetProvider | Pass terminalId to get_provider as id when openInGetProvider is true. When it is false, the provider is not in the provider directory yet (mostly unclassified rows), so cite the row’s own evidence instead. |
| results[].name, city, state, providerType, isSystem, beds, website | The provider. beds is null when unknown. |
| results[].evidence | strength (strong or some), titles (headcount and up to 3 verbatim titles) and website (tier 1–3, pages, share of the site, up to 3 URLs). |
| providersWithLine | Every provider with evidence for the line. |
| droppedUnclassified, droppedByFilter, droppedNoCity, droppedNoCard, droppedBySize | Why each provider read was not returned. droppedNoCity counts providers with no city on record, which a cities filter cannot match. |
| truncated | The scan cap stopped the sweep. Narrow by state. |
| asOf | When the nightly build last ran. |
- Unrecognised state — Unknown state: "<state>". with the list of valid codes.
- cities without state — cities needs state — city names repeat across states, so pass both.
- An empty list — No provider with evidence for the department matched. That means none recorded, not none exist — read the counters.
get_provider
One provider in depth: profile, health-system context including parent and largest members, technology relationships with verbatim evidence quotes, recent signals, and the clinical departments we hold evidence for. Each section says which kind of empty it is.
| Name | Type | Description |
|---|---|---|
| id required | string | Provider id or terminalId from search_providers. A name is not an id. |
{
"id": "100245871"
}One compact object, kept within a 20,000-character budget.
| provider | The profile, as a search_providers row. |
| system | The health system, its true member count, and up to 10 largest members. Omitted when the provider has no system. |
| edges | Up to 10 technology relationships, strongest evidence first. |
| signals | Up to 5 recent signals, one per source article. |
| serviceLines | Clinical departments, strongest evidence first — each with its label, strength, staff title headcount and website tier. An absent department is unknown, not missing. |
| meta | The resolved provider, a coverageNote for non-hospital providers, and truncated listing anything trimmed. |
- Unknown id — Provider not found: "<id>". — search first; names are not ids.
- A section is unavailable — { status: "unavailable", reason }: a source was down. Retry.
- A section is not tracked — { status: "not_tracked_for_this_provider_type", reason }: relationships and signals are linked for hospitals and health systems only today, so for other providers this is unknown, not zero.
get_provider_metrics
Every CMS metric a traced hospital’s page shows, about a hundred across financials, operations, cost centers, quality, patient satisfaction and network, each with what the page keeps one click down: the period it covers (a cost-report fiscal year with its end date, or a discharge window), the release, CMS’s own measure name, the source, the page’s colour band and the related rows (a star rating, averages, an excess readmission ratio, counts). Read from the same record the page is written from, in the page’s order. Only traced hospitals answer; an untraced hospital says so instead of serving its unverified import.
| Name | Type | Description |
|---|---|---|
| id optional | string | Provider id or terminalId from search_providers or get_provider. A name is not an id. Required unless segment is given; omit it with a segment to compare every traced hospital. |
| groups optional | string[] | Only these groups: financial, operations, costCenters, quality, satisfaction, network. Default all six. |
| limit optional | number | Rows returned in the page’s order, default 40, max 120. A traced hospital’s page shows about 100; pass groups to choose which rows count. With a segment and no id: hospitals returned, default 25, max 50. |
| segment optional | string | A revenue-cycle market segment id from list_segments (for example revenue-cycle-prior-authorization or clinical-documentation-integrity): read the hospital as that segment’s vendor. The reading names the rows that segment’s buyer cares about, why, and what this hospital’s numbers mean, each with its period. Reads from a vendor’s seat, so a provider or investor reader beside it is refused. |
| reader optional | string | Read the rows as this reader instead of the session’s own: vendor:sales, vendor:product, provider:finance, provider:clinical, provider:operations, provider:* (anyone at the hospital, its executives included: one headline row per group in a neutral voice) or investor:*. Default: the session’s persona department on its employer side. A hospital department with no reader of its own reads as provider:*, and a department at a company that sells to hospitals as vendor:sales, with readerSource “default”; an unknown or independent employer, or a payor, agency, media, law or consulting employer with no matching department, gets rows only. |
| series optional | boolean | Attach each row’s year-by-year history, oldest first, as {year, value, period, band}. Default true; pass false for a smaller answer. A row with no traced history carries no series. |
{
"id": "686695",
"groups": [
"financial",
"quality"
],
"limit": 60,
"reader": "vendor:sales"
}One compact object within a 46,000-character budget.
| provider | terminalId, CCN and the page’s name. |
| data | runId and writtenAt of the record, and the cost report it reads: fiscal year, period end, CMS report number and status. |
| reading | How to read the rows as the session’s reader (persona department crossed with employer side) or the reader passed: readAs and frame set the posture, never lists the misreadings to avoid, priorities are that reader’s rows in order, each with a one-sentence soWhat computed from this hospital’s own values and bands, askNext the questions to put back, and bands what a tier means. readerSource says whether the reader came from the session, the side’s default or the reader argument. Absent only when the employer side is unknown. |
| metrics | One row per metric: group, key, label, value (the number), display (the page’s string, unit included), tier (green, yellow, red or null), on a rated row rule (the direction and the exact green, yellow and red boundaries the band came from), period, released, measure, source, sourceUrl and related. Compare on period, never on a year: fiscal years end on different dates per hospital. A row whose history is traced also carries series, every year oldest first; a null value is a year CMS did not report, never zero. |
| withheld | Groups the key’s tier does not get, each with a reason. Never silent. |
| reading.segment | With segment: the segment read for, its why, the segment’s priority rows each with period and soWhat, notReported (rows this hospital does not file, which are not zero) and, for a segment no CMS number measures, noCmsSignal instead of rows. |
| hospitals | With segment and no id: every traced hospital, each with only the segment’s rows, ordered as meta.order says. meta.traced equals returned + unavailable + droppedByLimit + droppedByBudget; only traced hospitals are in it, so a ranking is a ranking of those, not of the market. |
| meta | totalMetrics and byGroup before the limit, truncated naming anything cut, seriesRows and seriesDropped (series cut to fit, from the end, before any row), and timingsMs. |
- Unknown id — Provider not found: "<id>". — search first; names are not ids.
- The hospital is not traced — { status: "not_traced", reason }: its page values are not yet checked against a public file, so none are served. get_provider still has its profile, system, vendors and news.
- A health system or a non-hospital provider — { status: "not_tracked_for_this_provider_type", reason }: CMS files per hospital; call this on the system’s member hospitals.
- The record is mid-refresh — { status: "unavailable", reason }: the rows and their parent are from different refreshes. Retry.
People
People with a current role at any hospital, health system, or published vendor, investor, payor or association, plus the alumni trail behind them.
get_persona
Research a role the way you would research a company. Real job titles with counts, how many people we track and at what kind of employer, duties quoted verbatim from real job postings, typical KPIs, buying role, and what people in the role say they care about — grounded in the census rather than in generic knowledge.
Ask it: “What does a CDI nurse actually do all day?”
| Name | Type | Description |
|---|---|---|
| query required | string | A role in plain words (“CDI nurse”, “VP of revenue cycle”) or an exact personaId. 1–200 characters. |
{
"query": "VP of revenue cycle"
}One object with status: "ok".
| personaId, label | The seniority-by-department group it resolved to. |
| roleDefinition | What the role is. |
| titleVariants[] | Real job titles held, with counts. |
| census | How many people we track, and at which employer types. |
| responsibilities[], kpis[], buyingRole | What the role does and is measured on. |
| jdQuotes | Quotes from real job postings, fenced as untrusted text. |
| aggregated | What people in the role say they care about. Withheld below an anonymity floor, and note says why. |
- No match — status: "no_route" with personaMenu[]. Pick the closest personaId and call again.
- More than one match — status: "ambiguous" with candidates[]. Call again with one of them; never guess.
search_people
Find people by name, employer or job title, everyone at one organization, or who runs it. People holding a leadership seat confirmed by our team or by the organization’s own leadership page come first, then everyone else by seniority. Every result holds a current role at a visible organization and lists the person’s other current roles. Pass at least one of query, organization_id, department or seat.
Ask it: “Who are the CEO, CFO and CNO at Albany Medical Center, and how do we know?”
| Name | Type | Description |
|---|---|---|
| query optional | string | Matched against name, employer and job title (and a confirmed seat’s title). 1–200 characters. |
| organization_id optional | string | Only people currently at this organization. A numeric terminal id works too. |
| vendor_id optional | string | Older name for organization_id. Pass one or the other. |
| department optional | enum | One of clinical, operations, general, finance, recruiting, customer_success, marketing, engineering, legal, data, product, sales, security, design, editorial, research. |
| role optional | enum | Seniority: c_suite, vp, director, or executive for all three. Does not count as a filter on its own. |
| limit optional | integer | Maximum results, 1–50. Default 20. |
| seat optional | string | Enterprise. Only confirmed holders of one seat: ceo, president, coo, cfo, cmo_clinical, cno, cio, cmio, cmo_marketing, chro, cqo, or another key a result returned (such as clo_legal). |
| confirmed_only optional | boolean | Enterprise. Only people holding a confirmed leadership seat. Default false. |
{
"query": "Albany Medical",
"role": "c_suite",
"limit": 10
}An array of people, one row per person: confirmed seat holders first, then the most senior role that matched.
| id | Person id — pass it to get_person. |
| name, jobTitle, organization | Who, and the role shown: the matched confirmed seat, otherwise their most senior matching role. |
| seats[] | Every confirmed seat the person holds: seatKey, title, chairTitle, tenure, organizationId, organization, confirmedBy (human = confirmed by our team, org_page = named on the organization’s own leadership page) and sourceUrl when recorded. tenure is permanent, interim or functional_head: the organization has no chief for that function and this person leads it at a lower rank, so title is their own title and chairTitle is the chair they occupy. Empty means none confirmed, not none held. |
| positionLevel | Seniority: CEO, President, C_Suite, VP, Director, … |
| functionalDepartment | Department of that role. |
| personaId | <level>_<department>, such as c_suite_ceo. An absent value means the role is unclassified, not junior. |
| imageUrl, linkedin | Photo and LinkedIn. A confirmed seat admits a person without a photo, so imageUrl can be absent. The profile headline is on get_person. |
| otherCurrentRoles[] | Up to 5 other visible roles, most senior first. |
| otherCurrentRolesHidden | Other roles at employers we cannot name. |
- No filter — Provide at least one of query, organization_id, department, or seat.
- organization_id and vendor_id disagree — An error asking for one of them.
- organization_id is not an id — An error saying to pass an id or terminal id, not a name.
- People held back — A second text block after the results counts people who matched but whose roles are at organizations not published here. Treat it as a coverage gap.
- Nobody confirmed (seat or confirmed_only) — A second text block counts people who match in employment records but have no confirmed seat, and seat holders at organizations not published here. Treat it as a coverage gap, never as “nobody holds that role”.
get_person
One profile in full: confirmed seats, title, employer, other current roles, past roles, headline, headshot and LinkedIn. Every id from search_people and search_alumni resolves here.
| Name | Type | Description |
|---|---|---|
| id required | string | Person id from search_people or search_alumni, or a slug. |
{
"id": "Xk29fPq7LmN3"
}One person object with the fields of a search_people row (seats[] is empty for a person with none; jobTitle is their most senior current role), plus:
| pastRoles[] | Up to 25 closed roles, newest departure first, with employer, title and dated start and end. |
| pastRolesWithheld | Past employers we cannot name. |
| upcomingEventAppearances | Up to five upcoming conferences the subject appears at, with roles. Absent when none are recorded; from news and LinkedIn, so absence is not evidence. |
- Not found — Person not found. Also returned for a person with no confirmed seat and either no current role at a visible organization or no profile photo.
search_executives
Retiring: use search_people, which now lists an organization’s confirmed leaders first. This returns the same answer as search_people with confirmed_only and stays for one release so connected assistants keep working. Enterprise. Pass at least one of query, organization_id or seat.
| Name | Type | Description |
|---|---|---|
| query optional | string | Matched against name, the seat’s organization, the seat title and job title. 1–200 characters. |
| organization_id optional | string | Only seats at this organization. |
| seat optional | string | One seat by key, as in search_people. |
| limit optional | integer | Maximum results, 1–50. Default 20. |
{
"query": "Albany Medical",
"limit": 10
}The rows search_people returns with confirmed_only: true.
- No filter — Provide at least one of query, organization_id, or seat.
search_alumni
The departures side of the graph. Anchor on a vendor, hospital, health system or any other published organization and get each person’s past role with a dated departure, plus every current role they hold, most senior first. Current employers are named when published and counted when not. Advisor, board and fractional seats are flagged, and when filters leave fewer than five people the closest misses come back too, each saying which filter it missed.
Ask it: “Who left Epic in the last year and where did they land?”
| Name | Type | Description |
|---|---|---|
| organization_id optional | string | An organization id from search_vendors, search_providers or resolve_hospital, or its numeric slug. A name is not an id. Pass this or vendor_id; one of the two is needed. |
| vendor_id optional | string | Deprecated alias of organization_id, kept for existing callers. |
| left_after optional | date | YYYY-MM-DD. Only people who left on or after this date. |
| role optional | enum | Seniority of the past role: c_suite, vp, director or executive. |
| department optional | string | Department of the past role, such as sales or engineering. |
| current_role optional | enum | Seniority now, as a floor: director keeps people who are a director or more senior today, whatever they were before. |
| limit optional | integer | Maximum people, 1–50. Default 25. |
{
"organization_id": "100000001",
"left_after": "2025-09-01",
"role": "executive"
}One object.
| organization | The anchor organization: id, name and type. |
| vendor | Deprecated copy of organization (id and name), kept for existing callers. |
| results[] | Newest departure first: the person, their pastRole at the organization with dates, and every currentRoles[] entry, with customer evidence for vendor employers. |
| totalMatched | Everyone with a past role at the organization, before filters. |
| _dropped* counters | Why matched people are missing, by reason. |
- No id passed — Pass organization_id — neither organization_id nor vendor_id was given.
- Unknown or unpublished organization — Organization not found: "<id>". — pass an id from search_vendors, search_providers or resolve_hospital.
- Undated departures — Excluded when left_after is set, and counted in the drop counters.
search_retirements
Executives who will retire or have retired, each with the last day, a confidence and who said so. Editor-confirmed retirements (a published story) come first; after them, people whose own LinkedIn profile says they have retired, marked low confidence. Every row is backed by a sourced claim, and the list grows as retirements are recorded.
Ask it: “Which health-system CEOs have announced they are retiring?”
| Name | Type | Description |
|---|---|---|
| status optional | enum | announced (last day still ahead), completed, or any. Default any. |
| seat optional | string | A seat such as ceo, cfo, cno or cio. |
| since optional | date | YYYY-MM-DD. Only last days on or after this date; rows with no last day are excluded. |
| min_confidence optional | enum | high returns editor-confirmed retirements only; low or any also returns LinkedIn self-reports. Default any. |
| limit optional | integer | Maximum people, 1–50. Default 25. |
{
"status": "announced",
"seat": "ceo"
}One object.
| results[] | Announced first (soonest last day first), then completed, editor-confirmed before self-reported: the person, organization, seat, lastDay, status, confidence, the reporting source and every claims entry. |
| retirementsRecorded | Every retirement on record. |
| droppedByFilter, droppedNoPhoto, droppedBySize | Why rows are missing; droppedBySize above zero means narrow the query or lower limit. |
| organizationHidden | People returned with the organization left blank because it is not published. |
| note | What the list does and does not cover. |
- Empty results — No matching retirements are recorded yet — not an error.
ask_leadership_editor
Hands a leadership question that needs several lookups to Lance Williams, Clinical Terminal’s AI leadership agent. It researches on its own with this server’s read-only tools plus a cited web search, and answers in prose with inline links and a sign-off line. Slower than one tool (up to about 90 seconds) and limited to a few questions per key per day; for a single fact, call search_people or get_person directly.
Ask it: “Who runs Atrium Health, and what changed in its C-suite this year?”
| Name | Type | Description |
|---|---|---|
| question required | string | The leadership question in plain words, 3–1,000 characters. |
| organization optional | string | Optional: the hospital, health system or vendor the question is about, by name. |
{
"question": "Who runs Atrium Health, and what changed in its C-suite this year?"
}Prose text.
| text | The answer with inline links, ending with the agent’s sign-off. Relay it as written; where it says something is not recorded yet, that is a coverage gap. |
- Paused or daily limit used — The agent says so; answer with search_people instead.
Signals
What the market did lately — news, funding, leadership changes.
search_signals
Recent market signals, newest first. Filter by text, by the entity a signal came from, by type, by source format, or by a detection window.
Ask it: “Show me every funding round in revenue cycle since May.”
| Name | Type | Description |
|---|---|---|
| query optional | string | Matched against title and summary. 1–200 characters. |
| entity_id optional | string | Only signals about this entity. |
| origin_entity_type optional | enum | vendor, provider (includes hospitals), hospital, investor or organization. |
| signal_type optional | string | An exact, case-sensitive category such as Funding / Investment or Leadership Change. |
| source_type optional | enum | The source format, such as news_article, press_release, linkedin_post, sec_filing or job_posting. |
| detected_after optional | date | ISO date or UTC datetime. On or after. |
| detected_before optional | date | ISO date or UTC datetime. On or before; a bare date includes the whole day. |
| sort optional | enum | importance or recency. Accepted, but results currently always come back newest first. Default recency. |
| distinct_articles optional | boolean | One row per article instead of one per article and linked entity. Default false. |
| limit optional | integer | Maximum rows, 1–50. Default 20. |
{
"signal_type": "Funding / Investment",
"detected_after": "2026-05-01",
"distinct_articles": true,
"limit": 50
}An array of signals, newest first.
| id, articleId | Row id, and the article it came from — group on articleId to count stories. |
| title, summary, sourceUrl | The story. |
| detectedAt | When we detected it. |
| topics[], categories[] | What it is about. Prefer these to signalType. |
| originEntityId, originEntityType | Which entity the row is about. |
| linkedEntityCount | How many entities the article names. |
| urgency, painPoint, salesAngle | Machine-written sales analysis. |
- Empty array — Nothing matched; there is no coverage note on this tool.
search_events
Published healthcare conferences and events: what is coming up, who is on an event’s roster (speaking, exhibiting, sponsoring, attending, hosting) and where a given person or organization is appearing. Events come from news and LinkedIn posts, not organiser registrations, and the list grows daily. An event with no roster means attendees are not recorded yet, never that nobody is going.
Ask it: “Which healthcare conferences are in Texas next quarter, and who is speaking?”
| Name | Type | Description |
|---|---|---|
| query optional | string | Words matched against the event title, host and location; all must appear. |
| starts_after optional | date | YYYY-MM-DD (Eastern time). Only events still running on or after this date. Default: today. |
| starts_before optional | date | YYYY-MM-DD. Only events starting on or before this date. |
| state optional | string | A US state as a 2-letter code or full name, e.g. TX or Texas. |
| city optional | string | A city name, e.g. Nashville. |
| format optional | enum | in_person, virtual or webinar (a single online session). |
| entity_id optional | string | A person or organization: its terminalId, or the id from search_people, search_vendors, get_person or a roster row. Returns the events it appears at. |
| role optional | enum | With entity_id: speaking, exhibiting, sponsoring, attending or hosting. |
| limit optional | integer | Maximum events, 1–25. Default 10. |
{
"state": "Texas",
"starts_before": "2027-01-31"
}One object.
| events[] | Soonest first (or your territory’s states first, see orderedBy): title, dates, location, format, host, link and, when attendees are recorded, a roster. A person carries an id only when get_person will open it; an organization carries an id and a followUp tool only when we hold its profile. speakingDetails is copied as recorded and may be the speaker’s own title rather than the talk. |
| orderedBy, territoryStates | How rows are sorted. |
| counters | Why events or ids are missing: filters, limit, response size, names without a profile. |
| coverage | How many upcoming events we hold, and how many have a roster. |
| snapshotBuiltAt, snapshotStale | When the event data was built; snapshotStale means it may be out of date. |
- Empty results — No matching event is recorded yet — not an error.
- Unknown state or id — An error naming the bad state or entity_id.
Graph
Who runs what. This is the only place hospitals are addressable as an anchor.
resolve_hospital
Turn a hospital name into the id the relationship lookup accepts. Resolution only, not a profile — it returns ranked candidates so you can pick the right St. Mary’s before you ask about edges.
| Name | Type | Description |
|---|---|---|
| query required | string | Hospital name or part of one. 1–200 characters. |
| state optional | string | A two-letter state code. Full names are not accepted here. |
| limit optional | integer | Maximum candidates, 1–20. Default 10. |
{
"query": "Massachusetts General",
"state": "MA",
"limit": 5
}An array of candidates, largest by staffed beds first.
| terminalId | The preferred get_relationships input. |
| id | Also accepted by get_relationships. |
| name, city, state | Which hospital this is. |
| hospitalType, staffedBeds | Type and size, to tell same-named hospitals apart. |
- Empty array — No hospital matched.
get_relationships
Company-to-company edges touching an entity, with provenance: source, evidence quote, evidence tier, a citable article and when it was first seen. The anchor can be a vendor or a hospital, which is what makes “what does this hospital run” answerable.
Ask it: “What technology does Mass General run?”
| Name | Type | Description |
|---|---|---|
| entity_id required | string | A vendor or hospital id, numeric slug or terminalId. Not a name. |
| edge_type optional | string | A relationship type such as CUSTOMER_OF or INVESTED_IN; case-insensitive. |
| limit optional | integer | Maximum edges, 1–200. Default 50. |
{
"entity_id": "100000001",
"edge_type": "CUSTOMER_OF",
"limit": 100
}An array of edges, active first. When some were withheld it is instead { relationships, withheld, note }.
| edgeType, relationshipLabel | The type to filter on, and its plain-English form. |
| label | The organization on the other end. |
| sourceId, targetId | terminalIds of the two ends. |
| evidenceSnippet, evidenceTier | The supporting quote, and its strength: 1 case study, 2 customer list, 3 logo or passing mention. |
| evidenceNamesLabel | Whether the quote actually names the other organization. |
| topArticleUrl | An article to cite. |
| firstSeenAt, lastConfirmedAt | When the edge was first and last seen. |
| corroborated, corroborationQuote | Independent corroboration, when checked. |
- Unknown id — Entity not found: "<id>". followed by how to pass an id: the id from search_vendors, a numeric slug, or a terminalId. A vendor name is not an id.
- Withheld edges — Counted in withheld: the other organization is unpublished and cannot be named. Say “at least” when counting.
Publishing
Tools that write. Everything above is read-only; the brand-kit tools below also save, to your own account only.
publish_page
Publish a page your assistant composed — a market analysis, a vendor write-up, an article — as a hosted page on clinicalterminal.com, laid out like a Desk story: headline, byline row with the publish date, and a right rail showing the companies and people it is about. Unlisted by default: only someone with the link can read it. A public page gets a keyword URL, a sitemap listing and a Markdown twin for other agents.
| Name | Type | Description |
|---|---|---|
| title required | string | Page title, 1–200 characters. Be specific. |
| body_html required | string | The page body only — what goes inside <body>, using the allowed tags and .ct-* classes. Up to 256 KB. |
| summary required | string | One sentence, up to 500 characters: the meta description and list subtitle. |
| citations optional | array | Up to 50 { label, url } sources, rendered in the rail’s Sources block. Default []. |
| source_tools optional | string[] | Which tools on this server produced the data. Default []. |
| visibility optional | enum | unlisted (unguessable link, not indexed) or public (indexable). Use public only when the user asks for a discoverable page. Default unlisted. |
| page_type optional | enum | comparison, analysis or article: the voice of the page and its default eyebrow. Every type uses the same layout. Guessed from the title when omitted. |
| artifact_id optional | string | Omit to create a page. Pass an existing id to update it in place at the same URL; visibility and page_type are then ignored. |
| entities optional | array | Up to 12 { type, terminal_id } entries — company (hospitals included) or person — naming what the page is about, by the terminal ids the research tools returned. They render as the Who This Touches rail, with a leadership card for the first company. On an update, omit to keep the rail; send a list to replace it. |
| draft optional | boolean | Rolling out: refused (nothing is published) until drafts are switched on for your account. true saves a private draft instead of publishing. Nobody else can see it; you read it at the returned preview_url and click Publish there, choosing link-only or public and whether to show your name. visibility is ignored on a draft. |
| base_revision optional | integer | Rolling out, with draft. On an update: the revision your assistant last read (get_my_page) or saved. If the page changed since, the save is refused instead of overwriting edits it has not seen. |
| charts optional | array | Up to 4 charts drawn into the page, the same forms send_to_desk takes: a vendor chart { key, kind: "vendor", vendor_tid, chart } (leader_origins, hired_from, departures, department_mix, seniority, capital_raised or customer_map; vendor_tid from search_vendors) or a hospital line chart. Place each with a paragraph that is exactly <p>[chart:key]</p>; it becomes a picture of the chart with its source note and a small table of its numbers. If any chart cannot be drawn, nothing is published and the reply says why. |
| chart_preview optional | boolean | true returns each chart’s headline and state, a line chart’s values by year, or the reason it cannot be drawn, and publishes nothing. Call it first, then write the page from those numbers. |
{
"title": "Texas acute-care hospitals: Q3 news activity",
"summary": "Twelve Texas systems drove most provider news this quarter.",
"body_html": "<h1>Texas acute-care activity</h1><p>Baylor Scott and White led signal volume.</p>",
"page_type": "analysis"
}One object.
| url | The page URL — give it to the user exactly as returned. |
| artifact_id | For later updates or revocation. |
| visibility | public or unlisted. |
| published_at, byte_size | When, and how large. |
| entities_dropped | Terminal ids from entities that are not on the rail (no match, or past the 12 cap). The publish still succeeds. |
| markdown_url | Public pages only: the agent-readable twin. |
| revision, note | On an update: the revision number and a note that readers see it within a minute. |
| status, preview_url | Drafts: status: "draft", the revision and the private preview_url where you read and publish it. A draft has no public url until you publish it. |
- Rejected HTML — Sending <!doctype>, <html>, <head>, <body>, <style>, <script> or <link>, an inline style or id, an event handler, or an image from an unapproved host is rejected with an error naming each offender. Nothing is silently stripped.
- Too large — body_html over 262,144 bytes. Cut sections rather than truncating.
- Daily quota — 20 publishes per key per UTC day, updates included, shared with publish_comparison.
- Storage quota — 200 live pages per key. Revoking one frees a slot.
- Update of an unknown id — No live page with that artifact_id is published under this account.
- Stale revision — The page changed since base_revision (it may have been published): read it again with get_my_page and reapply the edit.
- Drafts not enabled — A draft request while drafts are switched off: nothing is saved.
get_my_page
Rolling out: available once drafts are switched on for your account. Reads back one of your pages (a draft, or a live page) so an edit starts from the current text, not a stale copy.
| Name | Type | Description |
|---|---|---|
| artifact_id required | string | The id publish_page returned, or one from list_my_comparisons. |
| offset optional | integer | Byte offset of the body part to return. Long pages come in 40 KB parts: pass the previous reply’s next_offset until it is null. Default 0. |
| revision optional | integer | Required for a later part: the revision the first part returned. If the page changed between parts, the read is refused so two versions are never joined; read it again from offset 0. |
{
"artifact_id": "AbC123xyz"
}One object.
| title, summary, citations | The page as saved. |
| status, revision | draft or live, and the revision to pass as base_revision when saving the edit. |
| preview_url, url | Where you preview it; the public URL once live. |
| body_html, offset, next_offset, total_bytes | This part of the body and where the next part starts (null on the last part). |
- Unknown or unreadable id — No page with that id under this account, or it was revoked.
- Page changed between parts — The page was saved after your first part: discard the parts you have and read it again from offset 0.
publish_comparison
The same mechanics, shape and quota as publish_page, with guidance tuned for a vendor comparison, and page_type defaulting to comparison. A comparison assembled in a chat window dies in the scrollback; this is the version you can send to someone.
| Name | Type | Description |
|---|---|---|
| title required | string | Page title, 1–200 characters — “Waystar vs Adonis — denials, 400-bed IDN” beats “Vendor Comparison”. |
| body_html required | string | The page body only. Same allowed tags and .ct-* classes as publish_page. |
| summary required | string | One sentence, up to 500 characters. |
| citations optional | array | Up to 50 { label, url } sources. Default []. |
| source_tools optional | string[] | Which tools produced the data, such as get_vendor_dossier. Default []. |
| visibility optional | enum | unlisted or public. Default unlisted. |
| page_type optional | enum | comparison, analysis or article. Default comparison. |
| artifact_id optional | string | Pass an existing id to update in place. |
| entities optional | array | The companies and people the comparison covers, as { type, terminal_id }, up to 12 — the Who This Touches rail. |
{
"title": "Waystar vs Adonis: denials, 400-bed IDN",
"summary": "Waystar has broader confirmed hospital traction; Adonis is newer with regional references.",
"body_html": "<h1>Waystar vs Adonis</h1><p>Waystar leads on confirmed customers.</p>",
"citations": [
{
"label": "Waystar case study",
"url": "https://www.waystar.com/case-studies/"
}
],
"source_tools": [
"get_vendor_dossier",
"get_relationships"
]
}The same object as publish_page.
- Every publish_page error — Applies here too, against the same shared quota.
list_my_comparisons
Everything published under your key, newest first. This is your exposure audit — the one place to check what of yours is publicly readable.
| Name | Type | Description |
|---|---|---|
| limit optional | integer | How many to return, 1–50. Default 20. |
| include_revoked optional | boolean | Include revoked pages, flagged. Set false for live pages only. Default true. |
{
"limit": 10,
"include_revoked": false
}{ ok: true, artifacts }, newest first.
| artifact_id, title, summary | Which page. |
| url | The page URL; dead once revoked. |
| visibility | public (indexable) or unlisted (link only). |
| revoked | True once taken down. |
| published_at, byte_size | When, and how large. |
| view_count_approx | Approximate page loads — not analytics. |
revoke_comparison
Take a published page offline. The body is deleted rather than hidden and the URL stops working within about a minute. Permanent: republishing produces a new URL. Works for pages from either publish tool.
| Name | Type | Description |
|---|---|---|
| artifact_id required | string | The id a publish tool or list_my_comparisons returned. |
{
"artifact_id": "waystar-vs-adonis-denials-k3f9x2"
}{ ok: true, artifact_id }. One stored-page slot is freed.
- Unknown, not yours, or already revoked — The same answer for all three, so ids cannot be probed.
Brand kit
Clinical Terminal's own press kit, as data, and your own brand kits for any company: make one from the company's website, confirm it, and your designs come out in that brand.
get_brand_kit
The same kit that lives at clinicalterminal.com/pressrelease, returned as data. An assistant that calls it first uses the real marks; one that does not tends to invent a red. You can also keep your OWN brand kit for any company here: make one from the company’s website, then ask for it with company, and the designs come out in that brand with a “Data: Clinical Terminal” credit. Kits are private to the person who made them.
Ask it: “Design a LinkedIn banner in Clinical Terminal's brand”
| Name | Type | Description |
|---|---|---|
| company optional | string | Leave it out for the Clinical Terminal kit (unchanged). Any company’s name, id or terminalId returns your confirmed kit for it. Without one you get the Clinical Terminal kit plus a vendorKit.unavailable reason. |
| format optional | string | json (the default) or design_md: your kit as a DESIGN.md to save in your project. |
{}One object.
| assets[] | Every logo, wordmark, avatar, banner and cover: absolute url, kind, format, theme, pixel size and transparency. |
| colors[] | The brand colors as hex, each with its role. |
| fonts[] | The typefaces, their roles and where to get them. |
| usage, principles[] | Do and don’t rules, and one-line brand rules. |
| readMe | How to read the payload. Quote values exactly. |
create_brand_kit
Reads the company’s own website and saves a DRAFT kit: colors with a ground picked from the site, Google fonts, light and dark logos, corner radius and voice. It fixes what extraction gets wrong (the company name, button-text contrast, a logo-gradient highlight, accent contrast, a dark-ground logo, fonts not on Google Fonts) and lists each fix. The kit is private to you; re-running makes a new draft and never touches a confirmed kit. A free account can make 5 a month.
Ask it: “Make a brand kit for SmarterDx”
| Name | Type | Description |
|---|---|---|
| company required | string | Any company: its name, id (from search_vendors) or terminalId. |
| url optional | string | Staff only: read this page instead of the website on the company’s record. |
{
"company": "SmarterDx"
}One object.
| draft, draftId | The draft kit and the id to confirm. |
| fixes[] | Each fix as a before → after line to show the person. |
| next | The preview-then-confirm steps for the assistant to follow. |
confirm_brand_kit
Confirms your latest draft for a company, with any edits, so get_brand_kit serves it. Edits are checked: a 6-digit hex, button text at least 4.5:1 on the primary, a font on Google Fonts, a radius from 0 to 64 px. A rejection names the field. An older draft is refused.
Ask it: “Confirm it, but make the primary #00C8E2”
| Name | Type | Description |
|---|---|---|
| draft_id required | string | The draftId from create_brand_kit. |
| company required | string | The same company you passed to create_brand_kit. |
| edits optional | object | Changes to apply, e.g. { colors: { primary: '#00C8E2' } }: colors, fonts, radius and voice. |
{
"draft_id": "a1b2c3d4e5",
"company": "SmarterDx",
"edits": {
"colors": {
"primary": "#00C8E2"
}
}
}One object.
| kit | The confirmed kit. |
| next | Write DESIGN.md into the project and add it to CLAUDE.md. |
withdraw_brand_kit
Withdraws your confirmed kit for a company, so get_brand_kit stops serving it. It is archived with your reason, and never affects anyone else’s kit.
Ask it: “Withdraw my SmarterDx brand kit”
| Name | Type | Description |
|---|---|---|
| reason required | string | Why, in your words. |
| company required | string | The company: its name, id or terminalId. |
{
"reason": "We rebranded",
"company": "SmarterDx"
}One object.
| withdrawn | True when a confirmed kit was withdrawn. |
get_infographic_agent
The brief an assistant follows to turn Clinical Terminal research into a social story: a carousel of static frames, or a motion story you can download as MP4 or GIF. It covers which tools to research with, how numbers must be sourced, the logo rule, a storyboard you approve before anything is built, and how to build each style.
Ask it: “Make me a carousel on Abridge's momentum this month”
| Name | Type | Description |
|---|---|---|
| style optional | string | carousel (the default) or motion-story. Leave it out and the assistant asks. |
| topic optional | string | What the story is about. Leave it out and the assistant asks. Up to 500 characters. |
| vendor optional | string | A company’s name: brand the story in your confirmed kit for that company instead of Clinical Terminal’s. |
{
"style": "carousel",
"topic": "Abridge momentum this month"
}Plain text.
| the brief | Step-by-step instructions for the assistant. It does not research or build anything itself; the assistant then calls the other tools. |
Moonlight
Talk to your own Moonlight agent, the opportunity scout that emails you postings that fit your brief. Reads explain; writes change only your own brief and feedback.
set_moonlight_brief
Tells Moonlight, Clinical Terminal’s opportunity scout, what extra work or job you want. Moonlight then emails you postings that fit, on the cadence you choose. Two steps, never one: the first call reads your words and saves nothing; the second, after you confirm what it read, commits the brief and switches the emails on.
Ask it: “Have Moonlight look for clinical UX contract work for me, about 10 hours a week, remote.”
| Name | Type | Description |
|---|---|---|
| text optional | string | Step 1: what you want, in your own words (up to 2,000 characters). |
| token optional | string | Step 2: the token step 1 returned, once you have confirmed the targets. Expires after 15 minutes and works once. |
| cadence optional | string | Step 2: daily, weekly (default) or strong_only (only when something fits strongly). |
{
"text": "Clinical UX contract work, ~10 hrs/week, remote. Ambient AI vendors."
}One object.
| ok / result.status | Step 1: proposed with targets (craft, departments, engagementTypes, hoursPerWeek, geo, domains, seniority, exclusions, compFloor), missing andtoken; or parse_failed with one question to ask you. Step 2: confirmed with the brief’s version and cadence. |
- Proposal expired or already used — status: expired, invalid_token or no_proposal: start again from step 1.
- No Landscape subscription — ok: false, reason: not_entitled.
get_moonlight
Your Moonlight setup in one call: the brief (its text, chips and version), whether the emails are on and how often, the last nightly run and what it found, where mail goes (masked) and anything that would stop it, and how many cards were sent and still need your feedback.
Ask it: “Why haven't I heard from Moonlight?”
None — call it with an empty arguments object.
{}A short text answer, then one object.
| brief | text, targets, version, cadence; null before your first brief. |
| emails / lastRun / cards | On or off, cadence, masked address and blockedReason; the last run’s date and result; cards sent and awaiting feedback. |
- No Landscape subscription — reason: not_entitled.
explain_moonlight_matches
What Moonlight would send today for your saved brief, as a funnel: postings close to the brief, then recent, in your roles, (contract, for a contract-only brief), a close enough fit, and still open on the job board. When nothing would send, it names the step that emptied the list and the chip that would widen it. With try_edits it shows the funnel before and after those changes and saves nothing. Checks live job boards, so it takes a few seconds; limited to 40 calls per key per day.
Ask it: “Would taking full-time roles too get me more matches? Don't change anything yet.”
| Name | Type | Description |
|---|---|---|
| try_edits optional | array | Hypothetical chip edits, each {op, group, value}: op is add or remove; group is craft, departments, engagementTypes, hoursPerWeek, geo, domains, seniority, exclusions or compFloor. Not saved. |
| cadence optional | string | Hypothetical daily, weekly or strong_only. |
{
"try_edits": [
{
"op": "add",
"group": "engagementTypes",
"value": "full_time"
}
]
}A short text answer, then one object.
| preview, or before / after | stages (each step and how many postings are left), cards, sample (company and role) and bottleneck; with try_edits, both funnels plus the brief version to pass to edit_moonlight_brief. |
- No brief yet — reason: no_brief: write one with set_moonlight_brief.
- A value that cannot go in that group — reason: invalid_argument, naming the edit.
- Daily limit reached — The message says when it resets.
list_moonlight_cards
The opportunity cards Moonlight has emailed you, newest first: company, role, sent date, why it fit, the posting link and your feedback so far, each with the card_id give_moonlight_feedback needs.
Ask it: “What did Moonlight send me last week?”
| Name | Type | Description |
|---|---|---|
| limit optional | number | Cards per page, 1 to 20. Default 10. |
| cursor optional | string | The nextCursor from the previous page. |
| awaiting_feedback optional | boolean | Only cards you have not reacted to. |
{
"awaiting_feedback": true,
"limit": 5
}A short text answer, then one object.
| rows | cardId, companyName, title, sentAt, fitLine, jobUrl, outcome. |
| nextCursor / scanned / awaitingFeedback | Paging, and counts over your newest 200 cards (capped says when there are older ones). |
- No Landscape subscription — reason: not_entitled.
edit_moonlight_brief
Changes your saved brief straight away: add or remove chips (all apply or none do), set the cadence, or turn the emails off and on. Chip edits carry the brief version you last read, so a stale or repeated call is refused instead of applied twice. Returns the saved brief and today’s funnel. To rewrite the brief from scratch, use set_moonlight_brief.
Ask it: “OK, add full-time to my Moonlight brief, and only email me weekly.”
| Name | Type | Description |
|---|---|---|
| edits optional | array | Chip changes, each {op, group, value}, same groups as explain_moonlight_matches. |
| expected_version optional | number | The brief version you last read (get_moonlight). Required with edits. |
| cadence optional | string | daily, weekly or strong_only. |
| emails optional | string | on or off. The brief is kept either way. |
{
"edits": [
{
"op": "add",
"group": "engagementTypes",
"value": "full_time"
}
],
"expected_version": 3,
"cadence": "weekly"
}A short text answer, then one object.
| status | applied or unchanged, with the new version, targets, cadence, emailsEnabled and preview (the funnel after a chip edit). |
- The brief changed since you read it — reason: version_mismatch with the current version. Nothing was saved.
- A bad value, or an edit that empties the brief — reason: invalid_argument. Nothing was saved.
- No brief yet — reason: no_brief.
give_moonlight_feedback
Records your reaction to one card: interested, not_for_me, already_talked, mute_company (never this company again) or not_this_role (never this job title again). The last two add an exclusion chip to your brief. The first answer on a card is the one kept.
Ask it: “Not for me, and never send me Experian again.”
| Name | Type | Description |
|---|---|---|
| card_id required | string | From list_moonlight_cards. |
| action required | string | Your reaction, one of the five above. |
{
"card_id": "job_3f2a9c1d0b7e4a65",
"action": "mute_company"
}A short text answer, then one object.
| status | applied or outcome_already_set, with the outcome and any excluded value written to the brief. |
- Unknown card — reason: not_found.
Analysis
Ad-hoc SQL across the warehouse. These two are staff-only; every other tool on this page does not.
describe_schema
The column inventory for an allowed dataset, so an assistant can author correct SQL instead of guessing at table shapes.
| Name | Type | Description |
|---|---|---|
| dataset optional | enum | knowledge_graph, clinical_ai_design or analytics. Default knowledge_graph. |
| table optional | string | Restrict to one table. |
{
"dataset": "knowledge_graph",
"table": "providers"
}{ dataset, columns }, ordered by table then column position.
| columns[] | Each column’s table name, column name and data type. |
- Any failure — Never an error response: columns comes back empty with an error field. An unknown table returns an empty list with no error.
run_analysis_sql
A read-only query for analysis the fixed tools cannot express. One statement, writes rejected, a LIMIT 1000 added when you give none, and capped at 5 GiB scanned.
| Name | Type | Description |
|---|---|---|
| sql required | string | One SELECT or WITH query against the allowed datasets. |
| params optional | object | Named parameters, referenced as @name in the SQL. |
{
"sql": "SELECT provider_subtype, COUNT(*) AS n FROM `clinical-ai-design.knowledge_graph.providers` WHERE state = @st GROUP BY 1 ORDER BY n DESC",
"params": {
"st": "TX"
}
}{ ok, rows, bytesProcessed }.
- Any rejection — Never an error response: ok: false with an error naming the rule — multiple statements, anything but SELECT/WITH, a write keyword anywhere in the text, a dataset or project outside the allowlist, or a scan estimated over 5 GiB.
Limits & errors
These apply to every tool, before it runs. Tool-specific errors are listed on each entry above.
| Code | What it means |
|---|---|
| 401 | No authorization header was sent. |
| 403 | Invalid or revoked key, a lapsed Landscape subscription, or over the monthly quota. The response says which. |
| 429 | Per-minute rate limit. Retry after the interval the response names. |
The monthly cap is the control that actually binds. The per-minute limit is counted per server instance and the service scales out, so treat the stamped number as a floor rather than a promise.