Atlas Mind MCP Server

Atlas Mind

It's your mind, not someone else's.

Atlas Mind is the memory you share with your AI. Your notes stay plain files in your own git repository, served by an engine you can read line by line; your assistant reads and writes them over MCP, and every change says who made it, you or the AI, so you can see it, rewind it or undo it.

Create your Atlas → free, no card, no terminal · Live demo → the real viewer, in your browser · What is Atlas Mind?

Why Atlas

• Your files. Markdown with no syntax of its own, next to the PDFs, Word documents, HTML decks, images and JSON you already have, read as they are. All of it in a git repository that is yours: leave with a zip, open it anywhere.
• Your AI's memory. One address to give your assistant (Claude, ChatGPT, Claude Code, any MCP client): it signs you in over OAuth, asks for what it may do, then searches, reads, writes and files into your mind in every conversation.
• A memory you can trust. Every change is attributed, a person's or an AI's. See what an agent wrote and undo it in one click, ask what the mind said on a topic six months ago, find the notes nobody has looked at in a year and the ones that contradict each other.
• Yours from end to end. One Python program on the standard library: no database, no build step to run, nothing sent anywhere. Your notes never leave your server, and nothing you write trains anyone's model.

It is not a multi-tenant SaaS or a real-time collaborative editor: one focused mind per instance, fully yours, that can share slices with other Atlases.

Get started

Without a terminal, on alwaysdata's free plan (no card, hosted in France): seven steps, a few forms. From then on, Atlas updates from Settings › About.

On your computer:

uvx atlas-mind serve ~/my-mind     # try it once, nothing installed

uv tool install atlas-mind         # or: pipx install atlas-mind
atlas init ~/my-mind               # a new mind: atlas.toml, starter notes, AGENTS.md, git init
atlas serve ~/my-mind              # http://127.0.0.1:8765

On your server: atlas deploy ~/my-mind --target compose|systemd|fly writes the files and prints the guide (--target fly --wizard goes all the way). Atlas runs at the root of its address or under any sub-path (base_path = "/notes").

Requires Python 3.11+ and git. init never overwrites: it refuses a non-empty folder without --force and keeps the files already there.

Connect your AI

• claude.ai: on your Atlas's home, Add to Claude opens claude.ai with your Atlas filled in. Claude's free plan allows one connector: Atlas can be that one.

• Claude Code, for every project:

  claude mcp add --transport http --scope user atlas https://<your-atlas>/mcp
  

  then /mcp, pick atlas and sign in. Prompts read /atlas:weekly-review, a document attaches as @atlas:doc://notes/x.md. atlas init also installs a git hook so the commits Claude Code makes in a clone count as Claude's.

• Any other MCP client: give it https://<your-atlas>/mcp. Clients that can't do OAuth take a token instead, from Settings › AI assistants or atlas token create ~/my-mind --label my-agent.

Every access shows in Settings › AI assistants, with when it was last used, and can be revoked. An access can be limited to read, inbox and/or write: an agent that reads your mail never needs to edit a note. Your AGENTS.md reaches every client as its instructions, so an assistant follows your conventions the moment it connects.

What your AI can do

Thirty-four tools, in six groups:

It also serves prompts you start by name (weekly-review, triage-contradictions, new-doc, resume, and your own, one Markdown file each), and every note is a resource (doc://<path>) to attach to a conversation.

In hosts that show MCP Apps (Claude on the web and desktop, ChatGPT, VS Code), four answers come as views you work in: your recent work sessions to resume, the mind's weather folder by folder, what you thought then beside what you think now, and an agent's runs to play back commit by commit and cut. Other hosts get the same answer as text.

There is also a REST API v1 (Bearer tokens) with a published OpenAPI 3.1 spec.

In the viewer

• Read: GitHub-flavoured Markdown with a table of contents and [[wikilinks]], and PDF, .docx, HTML decks, images and JSON previewed inline. Full-text search, a command palette (Ctrl-K), backlinks, tags, pinned notes.
• The Mind: every note and tag a node, every folder a region, recent notes glowing, orphans dimmed. Above: a real mind, some 460 notes.
• Write and organise: create, edit, rename and move notes and whole folders (links follow), templates, drag and drop from your computer, a bin you can restore from or empty, a mind-wide task list.
• History and activity: every note versioned, with clean diffs and restore; a home with the week at a glance, a journal of who changed what with the AI's share, and an undo for what one writer did in one go.
• Inbox: what your agents staged, one card at a time: keep, snooze or bin, by click or keyboard.
• Team and sharing: accounts, groups, per-note access (private, shared, common), and public share links that can expire.
• Hive Mind: publish a folder as a node; another Atlas subscribes and keeps a read-only mirror, which its AI reads like any other note.
• Copies: a copy on your own GitHub, updated at every change; a copy on your computer (atlas sync); a full backup to download, and Restore on a new instance.
• Everywhere: an installable app on your phone (scan the code in Settings), offline reading, light and dark themes.
• Settings: a page with your account (name, password, recovery codes, 2FA, sessions), your AI's accesses, the team, the look (language, accent, the name and line under the logo) and About, with the update button.

Run it

• Security: admin and member accounts, per-note access, optional TOTP 2FA, recovery codes for a forgotten password, per-account lockout and per-IP rate limits, full CSRF defence.
• Network: by default Atlas only reaches out to read an MCP client's public metadata when you connect that client over OAuth ([oauth] client_metadata_fetch = false turns it off), and to ask PyPI for the latest version when an admin opens About.
• Configuration: one small atlas.toml: the name and line under the logo, the language, the port, accounts, the sub-path.
• Updates: from Settings › About on a pip install (alwaysdata), atlas update --target … for a deployment, uv tool upgrade atlas-mind on your computer. The engine and your notes update separately: upgrading never touches a note.
• Commands: atlas --help lists them: init, serve, build, deploy, update, user, token, share, inbox, import, export, sync, backup, restore, hook, doctor.

Documentation

The documentation is itself an Atlas mind: browse it in the live demo → (the guides/ and features/ folders).

Support

Atlas is free and open source. To help keep it going: sponsor it on GitHub. Bugs and ideas go to the issues.

Licence

AGPL v3, see LICENSE. The "Atlas" + "Mind" branding in the viewer and on the sign-in page is a fixed part of the project.