Atlas Mind is a self-hosted, git-native knowledge base where your notes stay plain files in your own private git repository, served by an engine you can read line by line — and an external brain you share with your AI over an open protocol. Own the data. Own the engine. Own the mind.
No account. No upload. The demo is the real viewer, running entirely in your browser.
The Mind — a live, force-directed knowledge graph of your notes.
Notion, and every hosted notes app, ask you to pour your thinking into their structure, on their servers, in their format. Leave, and you fight an exporter to get a tangle of files back. Your AI assistant reads your brain only by pulling it into their cloud.
Atlas Mind inverts that: a self-hosted Notion alternative, and a way to publish an Obsidian-style vault you actually own. Your knowledge never leaves your disk, and nothing you write trains anyone's model.
Everything ships in the box — a single Python server on the standard library, no database, no third-party network calls.
Markdown is first-class — rendered, linked, searched. But Atlas also previews standalone HTML decks, PDFs and Word .docx inline, converted in your browser. Nothing uploaded anywhere.
Atlas is the memory your assistant reads and enriches. It exposes an MCP endpoint with twenty-two tools — read, write, map, rewind and audit your mind — and atlas init scaffolds the conventions so an AI knows how to use it.
A self-hosted Markdown wiki with no database: one Python HTTP server plus a build step, entirely on the standard library. Accounts and share links are plain JSON on disk, fonts and libraries vendored, so it makes no outside calls.
Connect Atlas to Claude (or any MCP client) and your knowledge base stops being a passive archive: it becomes a persistent memory for Claude, an MCP server for your notes it can read, write, triage, map, rewind and audit, on your own machine, with nothing shipped to anyone's cloud. Twenty-four tools, six superpowers.
Ask in plain language and your AI searches the full text, opens any document, walks the tree and surfaces what you touched lately — across Markdown, HTML decks, PDFs and Word files alike.
search_docsread_doclist_treerecent_docsIt doesn't just answer — it files. Capture a decision, rewrite a section surgically, rename or refile a note with every [[backlink]] fixed automatically, archive what's stale (soft-deleted to trash, never erased).
create_docedit_docmove_docdelete_docYour AI traverses the wikilink graph the way you do: forward links, backlinks, a bird's-eye map of hubs and orphans, every note under a tag. It reasons about your mind's shape, not just its words.
get_linksget_backlinksget_mind_topologylist_by_tag“When did we decide this — and what was it before?” Your AI reads any past version of a note, diffs two points in time, finds a sentence you deleted months ago, traces a line back to the commit that wrote it, follows a moving changelog, and restores a version in one move — powered by the git history your mind already carries.
doc_historydoc_atdoc_diffsearch_historychangelogdoc_blamedoc_revertEvery change is attributed in the git history — who wrote it, when, and whether an AI made the edit. Your assistant reads the activity stream, surfaces stale notes untouched for months, and flags same-topic documents that may now contradict each other — then reviews them for you.
activitystalecontradictionsjudge_contradictionPoint your agents at the firehose: mail, alerts, a competitor's move, a CI webhook. They pre-sort it, and each item lands as a ready-to-file card in your own inbox, with a suggested folder, tags and the note it echoes. Keep, trash or snooze in a keystroke; Keep files it into your mind.
create_inbox_itemEvery tool runs locally and respects the same boundaries as the viewer — shared mirrors stay read-only, deletes are reversible, and not a byte leaves your disk. Your assistant gets a memory; you keep the keys.
Every change is a git commit, so the home page opens on a living record of how your knowledge grew — and who grew it. Each edit is attributed: its author, its date, and whether an AI made it.
A day-by-day feed of who changed what, with a badge on AI-authored edits. Click any entry to peek that document's history without leaving the page — and a "this week" digest sums up the pulse.
The people behind your mind drawn as a small constellation — a glanceable picture of who has been active lately, each star a contributor.
Two maintenance lenses: obsolescence — documents untouched for months — and contradictions — same-topic notes that may now disagree. A shortlist your AI can open and review.
If Atlas Mind disappeared tomorrow, your mind would still open in any text editor.
It's just a folder of files in your own private git repository. You are never holding someone else's format hostage.
Point the engine at a mind and it serves a fast single-page app. Upgrading the engine never touches your content; syncing content never touches the engine.
This repository: the Python server, the build step, and the single-page viewer. Read it line by line. Replace it. Your notes don't care.
Your notes, in their own private git repository: a content/ folder plus a small atlas.toml. State lives in a local .atlas/ — plain JSON, no database.
A wikilink [[graph]], full-text search, backlinks, tags, and a force-directed mind palace — everything a Zettelkasten, digital garden or PKM needs, built from your files.
Two Atlas instances can share a slice of their content and form a Hive Mind — a personal memory with sharing, not a real-time collaborative editor. It's pull-asymmetric: only the publisher needs to be reachable, the subscriber just pulls.
Pick a folder or a single document and publish it. You get a one-time atlas-node:… link carrying the origin URL and a read-only token (stored hashed, re-issue to rotate).
Paste the link on another instance. Atlas downloads the documents and keeps a read-only mirror under remotes/, re-syncing by SHA-256 diff on every pull.
Mirrored docs appear in the tree and as their own teal regions in The Mind. “Appropriate” forks an editable copy — and your AI reads them like any other note.
Tokens are scoped to the published path and revocable at any time. Nothing on your side is ever overwritten without a sync you control.
Installs a self-contained atlas command — the viewer ships inside the package, no separate download.
# install (uv fetches Python for you) — or: pipx install atlas-mind uv tool install atlas-mind # scaffold a new mind and serve it locally atlas init ~/my-mind atlas serve ~/my-mind # keep it current — upgrade the CLI, or update a deployment uv tool upgrade atlas-mind atlas update --target <fly|compose|systemd> # …or try it once, no install uvx atlas-mind serve ~/my-mind