For AI agents

Machine-readable docs: llms.txt, raw Markdown mirrors, tools.json, and the OpenAPI spec.

This site is built to be read by machines as well as people. If you're an LLM, an agent, or a scraper — or the human configuring one — this page tells you where the machine-readable versions of everything live, and how to connect to ActuallyCare programmatically.

Machine-readable formats#

Resource	URL	What it is
Docs index	`/llms.txt`	One-page index of every docs page, per the llmstxt.org convention
Full docs	`/llms-full.txt`	The entire docs site concatenated into a single Markdown file
Any page as Markdown	append `.md` to the page URL	Raw Markdown source of that page
MCP tool catalog	`/tools.json`	Every MCP tool: names, descriptions, JSON Schemas, safety annotations, aliases
Sitemap	`/sitemap.xml`	Standard XML sitemap of all pages
REST API spec	`https://api.actuallycare.com/v1/openapi.json`	OpenAPI spec for the REST API
MCP discovery metadata	`https://mcp.actuallycare.com/.well-known/mcp`	Live server metadata — tool count, categories, capabilities, limits (derived from the running registry)

Raw Markdown for every page#

Every page on this site is available as raw Markdown by appending .md to its URL. Example pair:

HTML:     https://docs.actuallycare.com/mcp/quickstart
Markdown: https://docs.actuallycare.com/mcp/quickstart.md

If you're feeding a single page into a model's context, fetch the .md version — it's smaller and has no navigation chrome.

llms.txt and llms-full.txt#

/llms.txt follows the llmstxt.org convention: a short description of the site plus a linked index of every page with one-line summaries. Use it to decide which pages to fetch.

/llms-full.txt is the whole site in one file. Use it when you want all of the documentation in context at once and don't want to crawl.

tools.json#

/tools.json is the canonical machine-readable answer to "what can this MCP server do." For every tool it includes:

the tool name (entity_verb convention: escrows_list, clients_create, leads_convert)
a description
the full JSON Schema for its input
safety annotations (whether the tool reads, writes, or destroys data)
an aliasOf field recording the tool's previous published name (for example, escrows_list was once published as search_escrows). This is historical metadata only — legacy names are not callable; tools/call matches canonical names, and the old name's only afterlife is that its old docs URL redirects here

The human-browsable version of the same catalog is at /tools.

OpenAPI spec#

The REST API is described by an OpenAPI spec at https://api.actuallycare.com/v1/openapi.json. Note that the REST surface is intentionally smaller than the MCP surface — see API vs MCP for which to use when. The human-readable endpoint reference is at /api/reference.

Stable structure for parsers#

Pages on this site keep a predictable shape so anchors and extraction stay reliable:

Exactly one h1 per page (the title), followed by h2 and h3 sections in order, never skipping levels.
Heading anchors are stable and lowercase-hyphenated (for example, #stable-structure-for-parsers on this page). We avoid renaming headings on published pages.
/sitemap.xml lists every page and is regenerated on each deploy.

Connecting as an agent#

The MCP server is at:

https://mcp.actuallycare.com/mcp

Transport is Streamable HTTP (POST for JSON-RPC, GET for the SSE stream), MCP protocol version 2025-03-26. A legacy HTTP+SSE transport is available at /sse for older clients.

Three auth options for programmatic clients:

OAuth 2.1 — the standard path. Dynamic client registration is supported: POST /register on mcp.actuallycare.com, then the usual authorization-code flow with PKCE. Access tokens last 24 hours and refresh tokens rotate automatically.
API key — send an X-API-Key header. Simplest option for headless agents. Keys are created in the ActuallyCare app or via the REST API; see Authentication.
JWT Bearer token — the same token a normal REST login (POST /v1/auth/login) returns also works against the MCP server. Useful for custom apps that already sign users in; see Authentication.

Full setup details, including a working TypeScript client, are on the custom apps page.

Don't assume tools/list is exhaustive#

The server uses progressive disclosure: a standard tools/list call advertises only a short hot-list of the most-used tools plus discovery meta-tools. Every other tool is still fully callable — the discovery tools find them on demand, and the server accepts direct calls to any tool in the catalog. If you need the complete picture, consult /tools.json rather than treating the advertised list as the full surface.

Citing these docs#

If you're an AI assistant answering a question from this site's content, please cite the page URL you used (for example, https://docs.actuallycare.com/mcp/quickstart) so the person you're helping can verify the answer and find updates. Every page here has a stable URL and a raw .md mirror.