A lightweight remote MCP server for Firm workspaces, designed to execute on Google Cloud Run.

This is a proof-of-concept showing how to give AI agents remote access to a Firm workspace backed by a Git repository. It wraps the tools from the firm_mcp crate and adds Git-based persistence and OAuth authentication on top. It is not intended for production use or larger teams. Think of it as a starting point you can adapt.

The server uses a Git repository as its data layer:

1. On startup, the server clones the repository and checks out a dedicated branch (default: mcp), merging in the latest from main.
2. Read operations (list, get, query, search, etc.) work against the local clone without modifying anything.
3. Write operations (add entity, write/replace/delete source) validate the change, apply it, and create an atomic commit pushed to the mcp branch.
4. After a period of inactivity, Cloud Run scales the instance to zero. The next request brings it back, starting from a fresh clone.

This means your agents can read and modify your workspace safely: changes land on a separate branch that you review and merge at your own pace.

• Single-user. There's no conflict handling. If the local state diverges from the remote, the server resets to match origin. Not suitable for concurrent use by multiple people.
• Stateless infrastructure. Auth codes and session state live in memory. When the instance scales to zero, they're lost. JWTs survive because they're self-contained.
• No dynamic client registration. The server uses pre-configured OAuth credentials. Your MCP client must be given the client ID and secret directly.
• Basic security hardening. Constant-time credential comparison, PKCE enforcement, redirect URI validation, JWT signing. Adequate for personal/small-team use, not vetted for high-security environments.

• A GCP project with billing enabled
• gcloud CLI installed and authenticated
• A GitHub repository containing your Firm workspace
• A GitHub personal access token with repo access

Copy .env.example to .env and fill in the values:

cp .env.example .env

For Claude, set the redirect URIs to:

ALLOWED_REDIRECT_URIS=https://claude.ai/api/mcp/auth_callback,https://claude.com/api/mcp/auth_callback

./deploy.sh

The script:

• Enables the Cloud Run, Cloud Build, and Secret Manager APIs
• Stores GITHUB_TOKEN and OAUTH_CLIENT_SECRET in Secret Manager
• Generates a JWT_SIGNING_KEY in Secret Manager if one doesn't exist
• Builds the container and deploys it to Cloud Run

On completion it prints the server URL and the credentials you'll need for your MCP client.

In Claude (or another MCP client that supports remote servers with OAuth), add a custom connector with:

The server implements the OAuth 2.0 authorization code flow with PKCE. On first connection, your client will complete the OAuth handshake automatically.

AGPL-3.0