Local MCP servers

MCP servers extend what your workers can do — each one is a small program askTheodor launches in the background, and the tools it exposes show up for your workers to use.

What you’ll learn

• What MCP is and how askTheodor runs a server.
• How to add one from a preset or as a custom server.
• How to test the connection and grant the tools to a worker.

What is MCP?

MCP (Model Context Protocol) is an open standard for giving AI tools. A server is a process that speaks the protocol and advertises tools; askTheodor is the client — it starts the server and lets your workers call its tools.

Add a server

1. Open Settings → MCP servers.
2. Presets (one click): a long row of + Name buttons covers Filesystem, Fetch, GitHub, SQLite, Brave Search, Gmail, Notion, Linear, Slack, Postgres, Stripe, and more. Click one, fill in any blank secret, and save.
3. Custom server: + Add custom server opens a blank form — a Command (npx, python3, docker, or an absolute path), Args one per line, and Env (KEY=VALUE) for tokens.
4. Click Test connection before saving — askTheodor starts the server and lists every tool it discovers, so you can confirm your command and tokens are right.

Grant the tools

Tools are namespaced <server-id>__<tool> and don’t activate per worker automatically. Open a worker’s Skills tab and add only the ones it needs.

Servers run real programs on your machine

An MCP server is an actual subprocess with whatever access you give it. Only add servers you trust, and scope their tokens narrowly (a read-only key, a limited IAM user).

Pair servers with the right worker

Give your research worker Brave Search and Fetch; give an ops worker GitHub and Linear. A focused toolset makes a worker faster and cheaper.

What this unlocks in practice

MCP is the difference between a worker that can write about your work and one that can reach into it. Three pairings that pay off immediately:

A researcher with Brave Search and Fetch. Now “research this competitor” means actually reading their live pages rather than recalling training data. The output changes character completely — it cites what’s on the site today, and it can tell you when something has changed since you last looked.

An ops worker with GitHub and Linear. “Summarise what shipped this week and which issues are still open” becomes a real query against real state. This is the point where the weekly-report Routine stops being a nice idea and starts being accurate.

A finance-shaped worker with SQLite or Postgres, read-only. Point it at a database with a read-only credential and ask questions in English. You get answers from your actual data, and the read-only key means the worst case is a slow query, not a modified table.

Scope the token, not just the worker

An MCP server runs with whatever access its token has. A read-only database user, a GitHub token limited to one repo, an API key with minimal scopes — do that work at the token level. The allowlist controls which worker can call the tool; only the token controls what the tool can actually do.

Recap

You added a local MCP server from a preset or by hand, tested it, and handed its tools to a specific worker. Next, skip the subprocess entirely and mount hosted servers straight from a directory.

📖 Reference in the docs: MCP servers

➡️ Next: The remote connector directory