Skip to content
StoryOS Docs

Use StoryOS with AI (MCP)

The StoryOS Model Context Protocol server exposes your workspace to AI agents — Claude first, ChatGPT-compatible by design. It’s a thin, stateless translator over the same /api/v1 the web app uses, so it can never drift from the product, and every response is scoped and validated server-side.

Two ways to connect

Section titled “Two ways to connect”
  • Local (stdio) — run the server as a subprocess of your MCP client (Claude Code, Claude Desktop). Best for a single user on their own machine.
  • Hosted (Streamable HTTP) — connect with just a URL and a token, no local process. Best for teammates and web clients (claude.ai / ChatGPT connectors, MCP Inspector). Cloud users can connect via a one-click OAuth connector instead of pasting a token.

Why it doesn’t hallucinate

Section titled “Why it doesn’t hallucinate”
  • Schema-first — describe_database gives exact api_names; the agent reads before it writes.
  • Validation-as-teacher — the API’s typed error is surfaced verbatim, so a wrong field or value comes back naming the problem and the model self-corrects in one turn.
  • Structured filters, not a query language — nothing to invent; the operator × type matrix ships in get_started.
  • Never invent ids — ids come only from search / list_* / a prior result; names and slugs are accepted and resolved server-side.

Because the MCP is a thin layer over the public API, an agent’s blast radius is exactly its token’s access grants — scope a PAT to one guest’s spaces and the agent can only touch those.

Next steps

Section titled “Next steps”