openserverless mcp

An MCP server for OpenServerless. It exposes the OpenServerless action tools (previously the @opencode-ai/plugin tools under trustable-app/tools/) over the Model Context Protocol so any MCP-capable agent can drive them.

Tools

The endpoint argument is either name (uses the v1 package) or package/name. Both segments must start with a letter and contain only letters, numbers, and hyphens. Use flat hyphenated names such as v1/employees-photo; underscores, spaces, nested routes, and forms such as v1/employees_photo or v1/employees/photo are invalid. Repeating action_new for a compatible existing endpoint is a successful check that leaves its files unchanged; incomplete paths and visibility conflicts are reported as MCP errors.

Secret values are never returned by these tools. Missing secrets and invalid endpoints are MCP errors (isError: true), so clients cannot mistake an incomplete binding for a successful tool call. action_add_secret remains as the single-endpoint compatibility tool; use secret_bind when several actions must share one user-configured credential.

Application .env and .env.production files are owned by Trustable's user-facing configuration flow. The MCP performs only value-free presence checks needed for binding validation and never creates, edits, imports, synchronizes, regenerates, or automatically populates those files. A missing variable must be added by the user through Trustable.

Application authentication uses Redis-backed opaque sessions. Create login, registration, me/session, every protected endpoint, and logout, then call auth_setup once with those complete endpoint sets. Store token-to-identity mappings under keys derived from ctx.REDIS_PREFIX with a bounded TTL; never use JWT or an application signing secret as a substitute.

Working directory

All path-based tools operate on paths relative to the process working directory: they read/write packages/<pkg>/<name>/... and .env. The server must therefore be launched with the user's app checkout as its CWD (this is how opencode launches type: "local" MCP servers).

Layout

src/
  index.ts            entrypoint — registers every tool over stdio
  lib.ts              shared helpers (endpoint parsing, connector injection, types)
  secrets.ts          secret name, .env, status, and atomic binding helpers
  tools/
    new.ts            action_new
    invoke.ts         action_invoke
    requirements.ts   action_requirements
    add-secret.ts     action_add_secret
    add-s3.ts         action_add_s3
    add-postgresql.ts action_add_postgresql
    add-redis.ts      action_add_redis
    add-milvus.ts     action_add_milvus
    add-mongodb.ts    action_add_mongodb
    secret-status.ts  secret_status
    secret-bind.ts    secret_bind
    secret-unbind.ts  secret_unbind
    auth-setup.ts     auth_setup

Each file under tools/ default-exports a Tool (defineTool({ name, config, handler })); index.ts imports them all and registers them in a loop. To add a tool, drop a file in tools/ and add it to the array in index.ts.

Run

Requires Node 18+ and ops on the PATH. TypeScript runs directly via tsx.

npm install
npm start          # tsx src/index.ts — speaks MCP over stdio

Wire into opencode

Add to the app's opencode.json:

{
  "mcp": {
    "openserverless": {
      "type": "local",
      "command": ["npx", "tsx", "/path/to/mcp/src/index.ts"],
      "enabled": true
    }
  }
}