MCP · read and write

Your memory, inside the AI you already use.

Tracebrain runs an MCP server, and a Claude plugin that brings it with the skills to use it. The assistant you already talk to can ask what you know about someone, what was said on a day, and what you owe — and write a task, a project or a fact about you back, without opening Tracebrain.

Start the 7-day trialSign in to connect

Connect it

No key anywhere, in either direction. You allow it once on a consent screen, the way you would any sign-in.

Claude — the plugin

The server and the skills together, in two commands.

claude plugin marketplace add tracebrain/mcp

claude plugin install tracebrain@tracebrain

  • Claude Code — the two commands above, then /mcp to sign in.
  • claude.ai and the desktop app — Customize → Plugins → add the marketplace tracebrain/mcp, then connect the server on the plugin's Connectors tab. That second step is real: a plugin cannot connect a remote server for you there.

It is a public repository, so you can read every line of it before you install it.

Everything else — the address

https://api.tracebrain.app/mcp

  • Claude, without the plugin — Settings → Connectors → Add custom connector, and paste the URL. Or open the dialog with it filled in. Or claude mcp add --transport http tracebrain https://api.tracebrain.app/mcp in Claude Code.
  • Cursor, Windsurf, ChatGPT, the Gemini CLI — add it as a remote HTTP MCP server with the same address.
  • VS Code — the same address in mcp.json, where the key is servers rather than mcpServers.

You will be asked to sign in to Tracebrain and allow access once. There is nothing to copy but the address above.

What it can do

10 tools read and 9 write, and they are the same ones the Tracebrain app itself uses. Nothing here is a general file interface: each one is a question or a change the product already knows how to make.

Reads

  • overviewWhat was decided lately, what is open, who you are working with
  • my_memoryYour own page: what the vault knows about you
  • find_peopleEveryone you talk to
  • personOne person's profile: the facts, the promises, the ideas
  • person_historyThe conversation itself, newest first
  • daysWhich days have a note
  • dayOne day: what was decided, promised, paid and planned
  • tasksWhat you owe, and what was promised to you
  • projectsThe work, with its people and its open tasks
  • projectOne project

Writes

  • add_taskAdd a promise to the task list
  • complete_taskTick one off, or put it back
  • create_projectStart a project
  • add_project_personPut somebody in a project
  • edit_memoryCorrect a line an automatic pass wrote
  • add_person_factAdd a fact to somebody's page
  • add_day_lineAdd a line to a day that is already written up
  • remember_about_meAdd something you told it to your own page
  • edit_my_memoryCorrect a line on your own page that a pass wrote

What the plugin adds

A tool list is not a product. Six skills say which tool to reach for and in what order, so you can ask for the thing you actually want.

  • briefBefore a call: who they are, what is open with them, what was last said
  • promisesWhat you owe and what is owed to you, and ticking things off
  • catch-upWhat happened over a stretch of days
  • project-statusOne project, its people and its open work
  • captureA promise you just made becomes a task on the right person
  • rememberOne durable line about you, onto your own page

And one hook, off until you switch it on: at the end of a session it asks whether anything you settled belongs on your own page. It never writes anything itself — it asks, and you see the line before it is written.

What it cannot do

The reassurance is not a promise about behaviour — it is that the tools to do these things do not exist.

  • Write into your message history, which is append-only
  • Delete anything: a note, a task or a line
  • Add a person, or create any note but a project it starts and your own page
  • Change a line you typed yourself
  • Reach another person's vault

Your conversation history is append-only by design, in Tracebrain and over MCP alike. The assistant can read it; nothing can rewrite it. How your data is handled.

FAQ

Questions people ask.

Which assistants can connect?

Any MCP client. For Claude there is a plugin, which brings the connector and the skills together in two commands. Everything else takes the address: Claude on the web and in the desktop app as a custom connector, Cursor, VS Code, Windsurf, ChatGPT and the Gemini CLI in their own settings.

What can it see?

Your vault and nothing else. Every call is signed in as you, and the vault it reads is the one your account owns. A business vault is reached only by a token you granted for that organisation on the consent screen.

Can it change my notes?

Only if you let it. One switch under Manage account decides whether assistants may change anything at all, and every write obeys it on the next call. With writing on it can add a task, tick one off, start a project, put somebody in a project, add a fact to somebody's page or to your own, and add a line to a day that has already been written up. Starting a project creates that project's note, the first task creates your task list and the first fact creates your own page; nothing else is ever created, and nothing is ever deleted. It cannot write into your message history, add a person, or change a line you typed yourself.

Can I give it read-only access?

Yes, with one switch under Manage account — not a choice your assistant makes for you. A permission scope is picked by the client at the consent screen, so read-only that way would mean trusting whatever it decided to ask for, and changing your mind would mean reconnecting. This you can turn off at any moment, and the next tool call obeys it.

What is the difference between the plugin and the address?

The same server either way. The plugin also installs six skills, so the assistant knows which tool to reach for and in what order, and one hook you can switch on that asks, once a session, whether anything you settled belongs on your own page. The address alone gives you the tools and nothing else, which is all some clients can take.

How do I disconnect it?

Remove the connector in your assistant. The permission is an ordinary OAuth grant in your account, so you can end it there too, and nothing is left holding a key.

Do I need an API key?

No. There is no key to copy, paste or store. You allow it once on a consent screen, the way you would any sign-in.