feega MCP

feega exposes a Model Context Protocol server so coding agents (Cursor, Claude, and others) can read and write a brand's canvas, posts and ad campaigns directly — with the same OAuth login as the CLI.

Your agent
   │  stdio (local)     →  bun run mcp  /  feega-mcp
   │  HTTPS (remote)    →  https://mcp.feega.app/mcp  + Bearer
   ▼
feega API  (/api/v1/*)

Quick start

Pick a transport: local stdio is simplest on your machine; remote HTTP is for cloud hosts.

Option A — Local stdio

Best for Cursor and other local agents.

1. Install Bun and clone the feega repo (or install the CLI binary so feega-mcp is on your PATH).
2. Authenticate once:

   feega login
   # once, in a terminal — the CLI and the local MCP server share one session

3. Add to Cursor MCP config (absolute path required when using Bun):

   {
     "mcpServers": {
       "feega": {
         "command": "bun",
         "args": ["run", "/ABS/PATH/to/feega/cli/mcp/stdio.ts"]
       }
     }
   }

   If the binary is on PATH after install:

   {
     "mcpServers": {
       "feega": { "command": "feega-mcp" }
     }
   }

4. Restart Cursor / reload MCP. Call `query` on `brands`, then work with a brand_id.

Session file (shared with the CLI): ~/.config/feega/session.json.

Option B — Remote HTTP

1. Confirm the server is up:

   curl -sS https://mcp.feega.app/health

   Expect a JSON health payload with ok: true and mcp: "/mcp".

2. Cursor MCP config:

   {
     "mcpServers": {
       "feega": {
         "url": "https://mcp.feega.app/mcp"
       }
     }
   }

3. Clients that speak MCP OAuth (opencode, Claude Code, Cursor, the Inspector…) authenticate on their own: on first connect they open feega.app, you approve, and the token is stored by the client. The initial 401 is the start of that handshake, not a crash.

Custom option — if your client cannot do OAuth, attach the token by hand: Authorization: Bearer <access_token>, using the Supabase access token that feega login saves in ~/.config/feega/session.json. Static API keys (feega_…) are not accepted by the remote server.

Option C — Local HTTP

cd cli
bun install
bun run mcp:http
# → http://localhost:8787/mcp
#    http://localhost:8787/health

Auth: Bearer or the local session file.

Auth rules

Protected resource metadata: GET /.well-known/oauth-protected-resource.

What to call first

1. query({ table: "brands", columns: ["id","slug","name"] }) — — confirm auth and see which brands/orgs this session can see
2. query({ table: "posts", columns: ["id","status","caption"], where: [...] }) — a brand's pending posts (add a brand_id filter once you have one)
3. Ids from list-derived reads accept short unambiguous prefixes; a `where` on `id` in `delete_row`/`update_row` needs the full id.

Twelve tools total — the CLI is narrower and brand-scoped, and does not expose the six generic ones below.

Tool areas

query reaches every table your org can see, as you: Postgres RLS returns exactly what you would see in the app. Always name columns, or the character cap drops whole rows to fit a long answer into a short one.

Full catalog, with parameters and CLI equivalents: tools.md.

Troubleshooting

Next

• Agent skill — install the feega skill so agents know when and how to call MCP or the CLI
• CLI — terminal commands when MCP is not available
• API reference — REST endpoints the CLI calls
• feega — source, tools map, and development notes