FAQ
What is FlowMCP exactly?
FlowMCP is a schema library + engine that normalises data sources into AI-callable tools. The library currently contains 373 production-ready schemas (v4) covering 291 data sources. The engine routes calls, validates inputs/outputs, and handles authentication. AI agents call FlowMCP; FlowMCP calls the underlying APIs.
Do I need an MCP-compatible client?
No. FlowMCP is CLI-first. The MCP server mode is optional. If you have a client that supports dynamic tool loading, MCP works — otherwise the CLI is the preferred entry point.
What does CLI-first mean in practice?
You run flowmcp call <schema>.<tool> '{...}' from your terminal, an LLM call, or a Node/Python script. Tools are loaded on demand — no need to keep 1608 tools in context.
Where do API keys live?
API keys live in ~/.flowmcp/.env, optionally with project overrides. The AI never sees a key — it only sees calls and responses. This is intentional: if an AI made direct REST calls, it would expose keys.
How do I add a new schema?
See the Specification and repos/flowmcp-schemas-private for the schema-creation guide. Validation rules have explicit IDs (e.g. RES001). PRs go to the schemas repo with tests.
Can I run FlowMCP offline?
Yes — for schemas that hit local resources (e.g. the gtfs-sqlite-toolkit add-on with a converted SQLite DB). For schemas calling remote APIs, you still need network access.
What is the spec version status?
FlowMCP currently supports the v4 specification (active) and v4.1 add-on extensions (e.g. gtfs-sqlite-toolkit). Older versions (v3, v2, v1.x) remain loadable for backward compatibility but receive no new schemas.
Why is FlowMCP "MCP" if it is CLI-first?
The name dates from the project’s start as an MCP server experiment. The substance shifted toward a CLI-first schema layer; the name persists for continuity with the MCP ecosystem.
Where can I report issues?
GitHub Issues per repo. Find the most relevant repository under github.com/FlowMCP and open an issue there. For schema-related questions, use flowmcp-schemas. For CLI bugs, use flowmcp-cli.
What is the v4 Self-Contained Skill Pattern?
Skills bring their own parameter reference along — schema data goes before workflow instructions. In a lab test, skills with full parameter/enum/example info hit 5/5 success against LLMs; skills with name + description only hit 0/5. The pattern is documented in the v4 specification.
Are FlowMCP schemas free to use?
Yes, MIT licensed. Some underlying APIs require keys (you supply); the data they return follows their own license — e.g. GTFS feeds under CC-BY 4.0 require attribution in your output. FlowMCP exposes that attribution in the response.