GuestGraphover MCP.

We are working toward hotels that know a returning guest as one person, whichever of their systems the guest passed through, from a profile they can ask why it believes what it does, and correct when it is wrong.

The endpoint

Three paths, and only one of them speaks the protocol. Add https://mcp.guestgraph.io/mcp to a client that takes a remote MCP server by URL.

• POST/mcpThe protocol. Streamable HTTP, one snapshot, no session.
• GET/healthWhether the server is up, and the model commit it answers from.
• GET/This page. A browser gets it; GET /mcp is a 405 on purpose.

Or run one against the same model over stdio, with no server at all:

npx --package github:companygraph/mcp-server companygraph-mcp ./model ./meta/core

What it answers

• list_typesEvery type the model's schemas declare. Use first, to learn which types exist before listing or describing one. No input. Returns types, each with type, name, tagline, owner (the type it nests under, or null) and count, the entities it holds.
• describe_schemaOne type's schema: its sections as written and its declared relations as data. Use to learn what an entity of the type may hold; for the whole vocabulary use describe_relations. Input: type. Returns url, the schema's own file, sections, and relations: owner, owns, references both ways, enums, joins and lists.
• describe_relationsWhat the schemas declare between types: references with form and cardinality, ownership, enums, joins, list kinds. Use for a diagram or audit; for one type's full schema use describe_schema. Optional type narrows every list, direction (declares, declared-to, both; needs type) one side of relations, via relations and enums. Returns those lists, with forms and reading explaining every term. Not paged.
• list_rulesThe rules the model is held to, from the CONVENTIONS.md its core vendors. Use to resolve a rule number that a schema or a refusal cites, such as R9. No input. Returns tagline, the file's url, and rules, each with rule, title and part; describe_rule gives one rule's text.
• describe_ruleOne rule as written. Input: rule, a number from list_rules such as R9, in either case. Returns rule, title, part, text and the file's url. An unknown number is refused with the known ones listed.
• list_checksThe checks the model's own gate runs, from the checker release this server was built with. Use to see which rules a script enforces and which are left to a reader. No input. Returns checks, each with name, rule and that rule's title, and ranBy. A list and no verdict: this server runs none of them.
• describe_errorsWhat a refused call looks like. A tool's output schema covers its answers only, so use this once, to branch on refusals or validate them. No input. Returns errors, every code with when it is raised and the JSON Schema of its details, and schema, the JSON Schema of a whole refusal: error with code, message, rule, details, beside model.
• list_entitiesThe entities of one type, by id. Use to browse a type; to find an entity by words or by name use search. Input: type, optional owner (an id) to keep one owner's entities, limit (default 50, at most 200) and cursor. Returns entities with id, type, name, tagline, owner, and page; follow page.nextCursor while page.hasMore.
• get_entityOne entity as structured data: fields, sections, tables and its references both ways. Use to reason over an entity; for its page as written use fetch. Input: id, or type and name; an ambiguous name is refused with candidate ids. Returns entity. Each reference list holds at most 50 edges; referenceCounts gives the totals and list_references the rest.
• list_referencesThe model's edges, filtered and paged. Use to inspect one entity's relations or one kind of reference without taking whole entities. Optional entity (an id), direction (out, in, both; needs entity), via, type (the far end's type, or either end's without entity), limit, cursor. Returns edges, each from, via, to, attrs, and page.
• find_evidenceEvery edge into one skill, grouped by the type of the page that drew it. Use to check a claimed skill against its evidence. Input: skill, an id or canonical name; optional limit, cursor. Returns skill, evidence and page; a profile's claim arrives via Skills.Skill with its level, an evidence row via Evidence.Skill. Attributes are verbatim.
• searchFind entities by words, substring or exact name. match: "words" needs every query word's stem in name, tagline, fields, sections or cells; "text" (default) is a case-insensitive substring over the same; "name" the exact canonical name. Optional type, owner, limit, cursor. Returns results with id, title, type, matched, words in words mode, and page: a listing, not a ranking.
• fetchOne entity's page as written. Use to quote or display the source; for structured fields and references use get_entity. Input: id, from search, list_entities or any reference. Returns id, title, type, url and text, the Markdown source. Takes no name: search with match "name" finds the id.

What it reads

A snapshot of guestgraph/mental-model@49662a2, parsed at build time against the core that instance vendored, core 0.40.0 with parser v0.45.0, served by companygraph-mcp-server v0.29.0. Every answer carries that commit, so an answer can be checked against the files it came from.