generated: '2026-07-19'
method: searched
status: published
source: https://flow-ai.com/docs/guides/mcp
summary: flowai-harness can expose its runtime tools as Model Context Protocol (MCP) servers, so an MCP-aware
  client (Claude, IDEs, agents) can call the data-agent tools defined in a Flow AI runtime. This is a
  capability of the harness rather than a single hosted endpoint operated by Flow AI - each host deploys
  its own server bound to its runtime and tenant.
server:
  name: flowai-harness
  transport:
  - stdio
  - streamable-http
  url: null
  docs:
  - https://flow-ai.com/docs/guides/mcp
  - https://flow-ai.com/docs/reference/mcp
notes:
- Runtime tools attached to agents (model-callable capabilities) are surfaced as MCP tools; the set of
  tools is defined per host runtime, not fixed by Flow AI.
- Transports are stdio (local) and Streamable HTTP (remote).
deployment:
  mode: none
  endpoint: https://flow-ai.com/docs/guides/mcp
  verified: probed
  probe: dead
  note: the endpoint this manifest claimed did not answer; recorded as none rather than deleted so the
    claim stays auditable
  checked: '2026-08-12'
  source: catalog MCP census