Skip to main content
Use this page to understand when MCP is the right integration path for Centaur and what the MCP surface exposes. For most external users, start with Client setup.

​
Overview

Use MCP when you want an agent or IDE client to call Centaur tools directly instead of working through raw REST requests.

​
When to use MCP

Choose MCP when:
  • your users work in Claude, ChatGPT, Cursor, Codex, or similar agent-capable clients
  • you want native tool calling instead of hand-crafted HTTP requests
  • you want agent workflows to stay aligned with the public Centaur skill repo
Choose REST when:
  • you need a conventional backend integration
  • you want complete control over HTTP calls and retries
  • your client does not support MCP

​
Endpoint

​
Authentication

Setup path:
  • Add the plain MCP URL above in a supported client
  • Let the client dynamically register itself when it needs a new OAuth client
  • Let the client complete the OAuth sign-in and consent flow
  • Reuse the issued access token for later MCP calls
Some OpenAI API-side MCP integrations may still require the calling application to supply an OAuth access token directly. This page covers interactive client flows first.

​
What Centaur exposes over MCP

Centaur exposes read-only data capability families over MCP:
  • events
  • messages
  • partner feed
  • generated channel narrative summaries
  • generated aggregate narrative summaries
  • positions
  • discovery
  • stats
Use Available data reads for exact read-family semantics and Contract limits for unsupported single-read requests and agent caveats.

​
Scope families

  • events.read
  • messages.read
  • feed.read
  • summaries.read
  • positions.read
  • directory.read
  • stats.read
A connection’s granted scopes are fixed when it is authorized, so a connection created before a scope existed does not gain it automatically. If a tool is missing from the inventory or returns a missing-permission error, disconnect and reconnect the Centaur MCP server once to refresh the grant. API keys are unaffected. The exact live tool and resource inventory is discoverable from the MCP server itself and may grow over time without changing the onboarding model on this page. Today, discovery and aggregate stats are separate read families, so agents can use list_traders and list_assets to resolve IDs before calling detail stats. The connected client can call any supported Centaur tool after setup, not only the example shown above.

​
Get started

Last modified on August 7, 2026