Regional endpoints
Your organization’s data lives in one Fluency region. Use the MCP URL for that region. A token issued for the other region is rejected. You will see If you are unsure which region your organization uses, sign in to Fluency as usual. Your Fluency admin or Fluency support can confirm the home region. Do not send the other region’s URL: org data does not cross regions.Who can call which tools
Access follows your Fluency organization role. Fluency reconciles that role with live WorkOS membership on every request, so a removed or demoted user loses MCP access without waiting for the token to expire.Connect a client
MCP clients discover AuthKit from the MCP URL. You do not paste a client id, client secret, or Fluency API key. On first use the client opens AuthKit; you sign in and consent to the organization you want the client to read.Cursor
- Open Cursor Settings → MCP.
- Add a new server with transport HTTP (sometimes labeled Streamable HTTP).
- Set the URL to your regional endpoint, for example
https://api.usefluency.com/mcp. - Save. Cursor fetches Fluency’s protected-resource metadata, registers itself with AuthKit, and prompts you to sign in.
- Complete AuthKit sign-in and organization consent in the browser window Cursor opens.
- Confirm the server lists Fluency tools such as
process_listandopportunity_list.
mcp.json entry looks like:
Claude
In Claude Desktop or Claude.ai connectors, add a custom MCP server and paste the same regional URL. Claude runs the OAuth discovery flow against AuthKit. Sign in as a Fluency user in the organization you want to query.ChatGPT and other MCP hosts
Any client that implements MCP Streamable HTTP and OAuth 2.1 (RFC 9728 protected-resource metadata) can connect the same way: give it the regional/mcp URL, complete AuthKit consent, then call tools. Fluency advertises:
- Resource:
https://<regional-api-host>/mcp - Authorization server: your AuthKit domain
- Bearer tokens in the
Authorizationheader only
What the tools return
Every tool is scoped to the organization you consented. You cannot pass another organization’s id; the server ignores org ids in tool arguments.
List tools are paginated. Defaults are conservative (typically 10 items; process executions default to 20, max 50). Prefer
*_list then *_get with a stable id from the list.
Software in MCP is Fluency Applications (the software catalog). Artifacts are Observe artifacts, not business-artifact or workflow-telemetry files.
Authentication details
- Sign-in is WorkOS AuthKit with OAuth 2.1. The MCP client is a third-party OAuth application; AuthKit shows a consent screen the first time you authorize it for an organization.
- Fluency product session cookies and dashboard JWTs are not accepted on
/mcp. You must complete the MCP OAuth flow even if you are already signed in to app.usefluency.com. - Internal Fluency machine tokens are a separate, Fluency-operated path. Customers do not mint those.
- Tokens are bound to one organization (
org_id) and, for user sign-in, to one region (organisation_region). Switching organization in AuthKit issues a different token. - If your Fluency role changes, the next MCP request asks you to re-authenticate. Reconnect the client rather than retrying with the old token.
Limits and behaviour
- Read-only. There is no write, delete, export-download, or settings tool.
- No Settings UI in Fluency for MCP. Adding the URL in the client is the entire setup.
- Tenant isolation. The organization on the token is the only tenant the tools can read.
- Rate limits.
/mcpuses the same API rate limits as other Fluency HTTPS routes. Back off on HTTP 429. - Errors. Failed tool calls return a generic tool error to the model. They do not include database connection strings or upstream stack traces.