Mu MCP Server

mu

A home for agents, tools and services

Overview

We're building services that operate as the building blocks for agents. Mail, chat, news, video, search, etc. Then we archive any data locally so it's all searchable. Services and the archive become tools for agents to use.

Using agents has also been a very synchronous experience. Communication here is async with everything going to one inbox, whether it's email, chat or sms, etc. Use it from the web, your phone, the command line, email, anywhere.

Features

What's included

• Services - 30+ real world services including news, mail, markets, video, etc accessible via API, CLI or the Web
• Agents - Define an agent by name, prompt and give it specific tools to use, then chat with it on the web, mail or xmpp
• Inbox - A single place to keep track of threads across channels. Use it on the web or via IMAP in an email client.

Services

The services available via the web, API or as tools, reachable via MCP, and command line

Open an issue to request a service.

Agents

An agent is a name, a prompt, and which of the services above it may reach. No code and no deployment — the services are the same ones in the table, so defining an agent is choosing a subset and saying what it is for.

Two come built in:

• Micro — general. Every service, for the questions that cross several of them: what happened today, what is moving, what is in my mail.
• Code — specific. A machine of its own and somewhere to put what it makes: shell and apps only. Describe an app and it writes it, runs it and hosts it.

Make your own at /agents: a name, a standing instruction, and the services it is allowed. Leaving the services empty gives it everything you can reach, which is a choice rather than the default.

Then talk to it wherever you already are:

mu ask "what is in my inbox?"                    # the default agent
mu ask --agent research "anything new this week?"

On the web it is /agent/<name>. By mail, write to you+name@ and that agent answers in the thread. Each keeps its own notes, so what it learns about your projects does not end up in the pool another agent answers from.

Agents run on the instance, using its model, so nothing needs a key of your own. mu agent is the other direction — see CLI.

Inbox

One place for what arrived, whichever channel carried it — mail, chat, sms, and the conversations you have had with your agents. It is a view over the record every client writes to, so a question you asked from the terminal this morning sits next to an email that came in overnight, in the same list.

It is not a second mailbox. service/mail is the mail server; /inbox is where things turn up. Reply to a thread and the agent answers in it.

Connect a normal mail client over IMAP and read the same threads from your phone.

Tools

If you want to use the tools with an existing agent.

Cursor, and clients with a config file. Create a token at /token:

{
  "mcpServers": {
    "mu": {
      "url": "https://micro.mu/mcp",
      "headers": {
        "Authorization": "Bearer ${env:MU_TOKEN}"
      }
    }
  }
}

Claude Desktop. Settings → Connectors → Add custom connector, and paste https://micro.mu/mcp. It registers itself, opens a browser and asks you to sign in — no token needed. Pasting the URL into claude_desktop_config.json will not work: that file only takes local command-line servers.

Anything else. It is JSON-RPC over HTTP POST:

curl -X POST https://micro.mu/mcp -H 'Content-Type: application/json' \
  -d '{"jsonrpc":"2.0","id":1,"method":"tools/list"}'

Scope the connection to the services you need:

https://micro.mu/mcp?tools=news,web,mail

See micro.mu/tools for all the tools.

Install

Quick install guide for self hosting (let us know if it's broken).

curl -fsSL https://raw.githubusercontent.com/micro/mu/main/install.sh | sh
mu --serve

Open http://localhost:8080. The first account you create is the admin.

Quite a few things need API keys, but here's some must haves.

Follow setup in CLI

mu setup        # pick an AI provider, paste a key
mu --serve

Everything else — mail, Google sign-in, Stripe for payments, etc — is optional.

The binary is a client, and by default it calls https://micro.mu — the instance this project runs live. Running your own? Point it there:

mu login https://your.host   # saves the address and a token
mu config get                # says which instance is in use, and why

Without that, mu news list on the machine you just installed calls the hosted instance rather than the one you are running. MU_URL and --url override per shell and per command.

Other ways to run it:

# Docker
git clone https://github.com/micro/mu && cd mu
docker compose up

# From source
git clone https://github.com/micro/mu
cd mu && go install
mu --serve

See the installation guide.

CLI

Every tool is a mu subcommand. The same binary runs the server (mu --serve) and the CLI.

mu news list                            # latest headlines
mu news search "ai safety"              # search news
mu web search "claude code"             # search the web
mu markets list --category stocks       # live prices
mu weather forecast --lat 51.5 --lon -0.12
mu docs list --collection notes         # your own documents
mu x402                                 # paying per call: config, and your key
mu help                                 # full tool list

Every tool in the table above is a command: the service, then the method. The underscore form works too, so mu news list and mu news_list are the same call.

The CLI is registry-driven — a tool added to the server automatically becomes a CLI command.

mu login                  # opens /token in your browser, paste the PAT back
mu config set token xxx   # or set it directly
export MU_TOKEN=xxx       # or use the environment

Run mu --help for the list — it reads the same catalogue the agent does.

These are tool calls: one command, one tool, no model involved and nothing to pay for unless the tool itself costs.

To talk to your agent instead, mu ask — it runs on the instance, so it needs your token and no model key of your own:

mu ask "what is in my inbox?"
mu ask --agent research "anything new this week?"

mu agent is the other direction and easy to reach for by mistake: it runs the agent here, on your machine, with your own model key, renting tools from an instance over x402 and paying per call. Same word in English, opposite ways round.

API

Every tool is also an HTTP endpoint, at /api/v1/<service>/<method>. The same catalogue the CLI and the agent read, so a tool added to the server is an endpoint without anybody writing a route.

curl https://micro.mu/api/v1/                      # the catalogue
curl "https://micro.mu/api/v1/news/list?limit=5"   # arguments in the query
curl -X POST https://micro.mu/api/v1/news/list \
  -H 'Content-Type: application/json' -d '{"limit":5}'

GET and POST mean the same thing and both work, because a REST API where reads are GETs is what every client library expects. What is not offered is a GET that changes something: a method the catalogue marks as writing needs a POST, so a link, a prefetch or a crawler cannot fire one.

Authenticate with a token from /token as Authorization: Bearer, or with an OAuth client. A priced endpoint answers 402 without one, which an x402 client pays per call with no account at all.

MCP clients connect at /mcp — same tools, same auth, same prices. The reference is at /api, and each service has its own page at /services/<name>.

Web

The same instance is a web app, and every page on it is a service answering.

• /home — what arrived, who is working on it, and what this instance knows right now, with a box at the top to ask.
• /inbox — the threads, whichever channel carried them.
• /agents — your roster, and where you make a new one.
• /services — the catalogue, and a page for each: /news, /weather, /markets, and the rest.

The service pages are not a separate product built over the API. They make the same calls an agent makes and render the result, which is why anything you can ask for you can also go and look at.

Configuration

What a call costs is data, not code: quota.json at the top of the repo is the one price list, and everything that charges or displays a price reads it. Drop a quota.json into the data directory to override any entry without rebuilding.

Some files are embedded in the binary, so editing means rebuilding:

• service/news/feeds.json — RSS news feeds
• service/chat/prompts.json — chat topics
• home/cards.json — home screen cards
• service/video/channels.json — YouTube channels
• service/places/locations.json — saved locations

See Install for every setting the code reads.

The rest lives in /admin/config on the server.

License

AGPL 3.0