Last updated: 2026-07-02

Production URL: https://api.outset.ai/mcp/

GitHub: Outset-AI/outset-mcp — client manifests (Claude Code & Cursor plugins, Gemini CLI extension, official MCP registry) and quick-connect snippets

Audience: developers building agents that need to connect to Outset. If you're comfortable with MCP and OAuth Dynamic Client Registration, the TL;DR below is enough to get you wired up — the rest of this page is for everyone else.

TL;DR (you already know MCP + DCR)

• MCP endpoint: POST <https://api.outset.ai/mcp/> — Streamable HTTP, JSON-RPC, stateless, single endpoint.
• Transport: mcp Python/TS SDK ≥ 1.27. No SSE, no WebSocket — every request stands on its own.
• Auth: OAuth 2.1 with PKCE (S256, mandatory). Access tokens 1 h, refresh tokens 30 d and rotated on use.
• AS metadata: https://api.outset.ai/.well-known/oauth-authorization-server — full RFC 8414 document with every endpoint URL, supported scopes, grant types, PKCE methods, and auth methods. Always fetch at runtime instead of hardcoding endpoint URLs.
• Client registration: CIMD (Client ID Metadata Documents) supported and preferred — use an https:// URL as your client_id, no registration round-trip. Open DCR per RFC 7591 still available. Loopback (http://127.0.0.1:*) or HTTPS redirect URIs only, either way.
• Default scopes: projects:read analytics:read studies:read. studies:write, analytics:write, and studies:launch (publish/unpublish + recruitment, may incur fees) are split out.
• RFC 8707: send resource=https://api.outset.ai/mcp/ on /authorize and /token. Tokens are audience-bound to the MCP server.
• No static API keys. The MCP endpoint is OAuth-only — there is no "paste a token" shortcut. Headless / no-browser clients use the RFC 8628 device authorization grant — see Headless environments below.

If you just want to try it: point npx @modelcontextprotocol/inspector at https://api.outset.ai/mcp/, click through the consent screen, and you're in.

What is the Outset MCP server?

The Model Context Protocol is an open spec from Anthropic for connecting LLM agents to tools and data. An MCP server exposes a list of tools (typed JSON-RPC methods) that any compliant MCP client — Claude Desktop, Cursor, MCP Inspector, your own agent built on the official Python/TypeScript SDK — can discover and call.

Outset's MCP server hosts the same study-creation, study-editing, and analytics tools that drive our internal AI co-editor and ATLAS analyst, exposed to outside agents. Concretely your agent can:

• Create and edit studies — projects, sections, questions, screeners, display logic, quotas, concepts, monadic configs.