Overview

Machine Memory’s remote Worker can act as a Model Context Protocol server. Agents connect with OAuth 2.1, then call tools to search and write project memories. Access is limited to the configured GitHub user ID and repository slug, not a shared bearer token.

Only GitHub OAuth is supported for now.

Who this is for

This is made for any users of any primitive that doesn’t allow or discourages persistent environments, such as Amp’s Orbs or chatbot-exclusive environments, where there is not a possibility to access such a tool.

Setup path

1. Infrastructure — deploy the Alchemy stack (Worker, D1, Vectorize, Workers AI).
2. Enable MCP — GitHub OAuth app + secrets, then redeploy.
3. Init — drop the managed MCP block into AGENTS.md (public script or machine-memory init --mcp).
4. Tools — connect a client to https://<worker-url>/mcp and use the tool surface.

Connect a client

Point an OAuth-capable MCP client at:

https://<worker-url>/mcp

Discovery endpoints:

• Authorization server: https://<worker-url>/.well-known/oauth-authorization-server
• Protected resource: https://<worker-url>/.well-known/oauth-protected-resource

On first connect the browser opens authorization: the client registers, GitHub identity is verified, and tokens carry mcp:read and mcp:write. After that, tools from Tools are available.

Troubleshooting

OAuth, cookie signing, sandbox credentials, and deploy gotchas: Troubleshooting: MCP.