About Mu

← All Docs

Tools for agents.

What Mu is

Mu gives an agent the everyday internet as tools: news, web search, mail, markets, weather, video, places, images, files, events, contacts. 67 of them, behind one MCP endpoint.

Point an MCP client at /mcp and everything is available. Browse what’s there at /tools.

Real tools, not wrappers

Most things offering an agent tools are a thin layer over somebody else’s API — reselling access, no margin, nothing to defend. Mu runs the things it exposes.

mail_inbox reads a real inbox, served by an SMTP server with DKIM this instance runs. web_search queries an index it maintains. db_create writes to real storage. apps_run executes in a sandbox it owns.

That is the whole idea. An agent is only as useful as what it can actually reach, and most of what it can reach today is either a chat box with no hands or a proxy to an API somebody else controls.

Two doors, one set of tools

Every capability is a service, and a service is reachable through either door. Nothing is built twice.

An agent calls the MCP endpoint at /mcp. It gets a tool list and uses what is there. That is the only way in — there is no second protocol to choose between, and nothing to integrate per tool.

A person signs in and gets the home screen: cards for each service at a glance — headlines, prices, weather, unread mail — with the agent inline to act on what they are looking at. Apps, saved items, wallet, settings, all of it.

Same services underneath. When a new one registers it appears in both places at once, with no extra wiring — the agent gets a new tool, the app gets a new card. That is why the tools are worth trusting: they are used by a person every day before any agent calls them.

Why an agent, not a developer

The previous generation of this idea sold APIs to developers. A developer has to find you, choose you, integrate you, and keep choosing you — high friction, high churn, competing on price with whoever is cheapest.

An agent does not shop. It is handed a tool list and uses what is there. That inverts the funnel: you are not persuading anyone to integrate, you are present at the moment of need.

Connecting

  1. Add the endpoint to your MCP client: https://micro.mu/mcp. Nothing else goes in the config.
  2. Sign in when it asks. The first call returns 401 pointing at this instance’s authorization server; a client that speaks MCP authorization walks you through sign-in and keeps the token itself. One that doesn’t takes a Personal Access Token from /token.
  3. Call anything. Reading this instance’s own content is included. Calls that cost money to run — a model call, a paid third party — draw credits.

It is the same account either way: sign into the app and your agent’s calls draw on the same balance.

What we don’t do

  • No ads — we don’t sell your attention
  • No tracking — we don’t profile you
  • No algorithmic ranking — chronological, transparent
  • No infinite scroll — there’s always an end
  • No push notifications — you come when you want

Technology

One Go binary, built on Go Micro. Every capability is a go-micro service; the assistant is a go-micro agent that calls them; /mcp is served by go-micro’s MCP gateway. No external infrastructure.

Supports Anthropic Claude, Atlas Cloud (DeepSeek, Qwen), or local models via any OpenAI-compatible API (Ollama, vLLM, llama.cpp).

Run an instance and you are the operator: the tools are yours, and anyone paying to call them pays you. AGPL-3.0.

Where to go next

  • Tools — the catalogue, with prices
  • MCP — protocol detail and playground
  • CLI — every tool as a mu subcommand
  • Installation — run your own
  • Architecture — how the services fit together

Try it at micro.mu — or self-host from github.com/micro/mu