Grantd

OAuth-for-agents: a dead-simple, MCP-native OAuth token broker that lets AI agents securely act on a user's behalf across third-party APIs (Gmail, Slack, GitHub, Notion…). Tokens are vaulted server-side and never touch the LLM.

Stack: TypeScript · Hono · Postgres · generic OAuth2 driven by a declarative provider registry · envelope encryption (AES-256-GCM, key-versioned) · Postgres advisory locks for refresh concurrency (no Redis). Security model and how to report issues: SECURITY.md.

Security at a glance

Grantd vaults users' OAuth tokens, so security is the product, not a feature. The fundamentals:

• Tokens encrypted at rest — AES-256-GCM envelope encryption, fresh IV per record, key-versioned, and fail-closed (a missing key refuses to write; it never silently stores plaintext).

• Tokens never reach the LLM or the caller. The proxy injects the access token at the network boundary and returns only the provider's response — the raw token is never serialized back.

• API keys hashed at rest (pbkdf2-sha256, peppered); the raw sk_ key is shown once.

• No SQL injection — every query is a parameterized postgres.js tagged template.

• Tenant isolation on every query by environment id; Row-Level Security enabled on all tables so a Postgres REST layer (e.g. Supabase/PostgREST) can't read the vault.

• 256-bit random OAuth state + session tokens; PKCE (S256) where the provider supports it.

• Durable, Postgres-backed rate limiting — per secret key on auth routes, per IP on public routes.

This is open-source infrastructure you run yourself. Self-hosting responsibilities (key custody, TLS, an edge/WAF for volumetric DoS) and known limitations are documented honestly in SECURITY.md. If you intend to run a hosted, multi-tenant deployment that holds other people's tokens, work through HOSTED-CHECKLIST.md first.

Related MCP server: Google Workspace MCP

Providers

Adding a provider is data, not code (see src/providers.ts). Help verifying Slack/Notion is welcome.

Quick start (dev)

npm install
npm run keygen            # prints an ENCRYPTION_KEYRING + API_KEY_SALT
cp .env.example .env      # then paste keygen output + your DATABASE_URL
npm run migrate           # apply SQL migrations
npm run dev               # start the broker on :8787

MCP server

Exposes the broker to an AI agent (Claude Code / Cursor / Claude Desktop) as tools: list_providers, check_connection, create_connect_link, and call_provider. The auth-gating wedge: if the agent calls call_provider for a user who isn't connected, it returns an authorization_required result with a connect link instead of failing.

The MCP server is published to npm as grantd-mcp, so you don't need to clone this repo to use it — just point an MCP client at it (you do need a running broker and a secret key). For local development: npm run mcp (needs the broker running); smoke test: npm run mcp:test.

Add to Claude Code:

claude mcp add grantd \
  --env GRANTD_API_KEY=sk_... \
  --env GRANTD_BASE_URL=https://your-broker.example.com \
  --env GRANTD_END_USER=user-123 \
  -- npx -y grantd-mcp

Or in a Cursor / Claude Desktop mcpServers config:

{
  "mcpServers": {
    "grantd": {
      "command": "npx",
      "args": ["-y", "grantd-mcp"],
      "env": {
        "GRANTD_API_KEY": "sk_...",
        "GRANTD_BASE_URL": "https://your-broker.example.com",
        "GRANTD_END_USER": "user-123"
      }
    }
  }
}

Layout

migrations/   SQL schema (Supabase/Postgres compatible)
scripts/      keygen + migrate helpers
src/
  config.ts   env loading + validation + keyring parse
  crypto.ts   envelope encryption + API-key hashing
  db.ts       postgres.js client
  providers.ts declarative provider registry (Google, GitHub, Slack, Notion)
  oauth.ts    generic OAuth2 client (authorize / exchange / refresh) driven by the registry
  server.ts   Hono app + routes (WIP)