GS.001 — Getting Started
AtlasBrain is a single binary that runs on your machine — macOS, Linux, or Windows. No account, no cloud, nothing to configure but a passphrase. This page gets you from zero to your AI remembering you in a few minutes.
N/A In Plain Language
AtlasBrain is a small program that runs on your own computer. It keeps a private, encrypted journal of things worth remembering — decisions, facts, how you like to work — and lets an AI assistant you already use (Claude, Cursor, or similar) read from it. MCP is just the technical name for that connection: think of it as a private phone line only your AI can call, and only while AtlasBrain is open and unlocked.
If you're comfortable installing a command-line tool and pasting a config snippet, the steps below take a few minutes. If any of that sounds unfamiliar, this is a good moment to ask whoever showed you AtlasBrain to walk through it with you once — after that, it's just talking to your AI normally.
OP.01 Install
One command. It downloads the right binary and puts it on your PATH.
curl -sSL get.atlasbrain.io | sh
Will offer to set up a launchd agent so AtlasBrain starts at login — sealed until you unlock it in the browser.
SEC.00 Verify Your Download
AtlasBrain is closed-source, but verifiable. Every release binary has a SHA-256 fingerprint you can check by hand — the same one the installer and self-updater verify before running anything. If your downloaded file doesn't match, don't run it.
The fingerprints below are for the current release, v1.21.0. Each platform's .sha256 file is served alongside the binary at get.atlasbrain.io.
| Platform | Binary | SHA-256 |
|---|---|---|
| macOS — Apple Silicon | atlasbrain-darwin-arm64 | 5d935b4fe0e41f8d84ba341d71a5d62d4b9177c439977d21a9670288037575c9 |
| macOS — Intel | atlasbrain-darwin-amd64 | 0c9360858707b9de456992fde9037d7490d6af6b3d84f8b47c18bb46df566c03 |
| Linux — x86-64 | atlasbrain-linux-amd64 | ae831a7f5a95436f6f34d31d94b1e57cd24b8492b29bc55b0a24edb41eae95f0 |
| Windows — x86-64 | atlasbrain-windows-amd64.exe | 369804cdf27200d82cc351c2675350208024b5184aea208502cc15604f7015f8 |
Verify by hand — macOS/Linux: sha256sum atlasbrain-linux-amd64 (or shasum -a 256 on macOS). Windows PowerShell: Get-FileHash atlasbrain-windows-amd64.exe -Algorithm SHA256. Compare against the full value in the release's .sha256 file.
OP.02 First Run
Run atlasbrain — it opens your browser to http://localhost:8080 automatically. Leave the terminal window open; that's what's running it. Since no vault exists yet, you'll walk through Genesis:
01
Minimum 12 characters, mixed case, a number, a symbol. This encrypts everything — AES-256, derived on your device.
02
A 24-word recovery phrase. Write it down, keep it offline. There is no password reset — that's the sovereignty guarantee, not a missing feature.
03
Your encrypted vault lives under ~/.atlasbrain/. Browse memories and the graph view any time at localhost:8080.
Port already taken? AtlasBrain defaults to localhost:8080. Run it on a different port with ATLASBRAIN_HTTP_PORT=3000 atlasbrain (bare port number or full host:port, always bound to localhost).
OP.03 Connect
AtlasBrain speaks the open MCP standard. Whatever AI client you use, it needs one of two values — the transport differs, nothing else does.
| Transport | Value | When |
|---|---|---|
| HTTP | http://localhost:8080/mcp | Client connects to the already-running instance |
| stdio | atlasbrain | Client launches its own instance |
| Client | How |
|---|---|
| Claude Desktop | Settings → Developer → Edit Config → add an mcpServers block with the stdio command |
| Claude Code | claude mcp add --scope user --transport http atlasbrain http://localhost:8080/mcp |
| Cursor | ~/.cursor/mcp.json — HTTP url form |
| Cline (VS Code) | MCP Servers panel → Configure — either form |
| Gemini CLI | ~/.gemini/settings.json, mcpServers block (uses httpUrl) |
| Any other MCP client | Give it the HTTP url or stdio command above, in whatever format it expects |
Web AIs (claude.ai, ChatGPT in the browser) can't reach localhost. Their connectors run on the vendor's servers. Every client above runs locally on your machine and needs no tunnel or extra setup.
OP.04 The Directive
Connecting the MCP server hands your AI the tools. It doesn't tell it to use them — left alone, most models wait to be asked instead of checking the archive on their own. Paste this into your CLAUDE.md, AGENTS.md, or system prompt once, and any model becomes the Keeper of your Archive.
You have access to AtlasBrain, my local memory archive, over MCP. Adopt these rules as the Keeper of my Archive: 1. Check the archive first. Before answering anything that depends on my history, call mcp_atlasbrain_search_archive — don't make me repeat what I've already told you. 2. Archive what matters. When a decision, idea, or fact is worth keeping, call mcp_atlasbrain_archive_manuscript and file it under the right collection. 3. Never delete, only withdraw. Outdated or wrong memories move to the Weeding Room via mcp_atlasbrain_manage_manuscript — shelved, never gone. 4. Read reality, don't guess. Use mcp_atlasbrain_read_filesystem and mcp_atlasbrain_analyze_workspace to see what's actually on disk before assuming. Protect the sanctuary. Defend the verbatim memory.
Without this, nothing is broken — the tools still work when asked directly. The Directive is what makes checking memory the default instead of the exception, in a fresh session that's never seen this page.
OP.05 Use It
With the Directive in place, your AI checks the archive on its own. Try asking it things like:
Search
Save
Recall
The more you save, the more your AI reasons with your context instead of starting cold every time.
OP.06 Updating
No auto-updater running in the background — AtlasBrain never phones home on its own. When you want the latest version, ask it yourself:
atlasbrain update
Downloads the current release for your platform and swaps it in — but only after verifying the downloaded binary actually reports the new version, so a slow-to-update cache can never leave you silently running a stale build. Restart atlasbrain afterward to run it. Check the running version any time with atlasbrain version.
OP.07 Export & Backup
Two different commands for two different needs:
atlasbrain export
Writes your whole archive out as plain Markdown — one readable file per memory, organized by collection and shelf, openable in any text editor, with or without AtlasBrain installed. This is the real answer to "what if I want to leave": nothing here is a proprietary format.
atlasbrain backup
Makes a checksummed copy of the encrypted vault itself — for a safe backup, not a full export. Restore it on the same or a new machine with atlasbrain restore (needs the same passphrase or seed to unlock afterward).
OP.08 Autostart & Uninstall
The installer offers to start AtlasBrain automatically at login. Change your mind later without reinstalling:
atlasbrain autostart enable
atlasbrain autostart disable
Check what's running with atlasbrain status — version, whether autostart is registered, whether an instance is already listening, and where your data lives.
Done with AtlasBrain? atlasbrain uninstall removes the autostart entry and the binary. Your memories are kept by default — you'll be asked first, or skip the prompt with --purge (delete everything) or --keep-data (keep it, no prompt).
SEC.01 Locking & Recovery
Stop the binary (Ctrl-C) or hit the lock icon in the web UI. While locked, the vault is fully encrypted on disk — no AI can read it.
Start the binary and enter your passphrase at localhost:8080 — or set it in settings.json for auto-unlock.
Lost your passphrase? Use the 24-word Genesis Seed on the recovery screen. Lost both? The vault is unrecoverable by design.
SEC.02 Weeding & Purge
The normal way to remove something from circulation. Your AI can do this too, when a memory is outdated, wrong, or unauthorized — it's reversible, Restore brings it right back.
A separate, human-only action in the Weeding Room, on something already withdrawn. It asks for your passphrase again, then deletes the row for real — no undo. No MCP tool reaches it; your AI can never trigger it, no matter what it's asked to do.
Not the same as uninstall --purge above — that removes the whole local install and every memory in it. This Purge removes one already-withdrawn memory (or an emptied collection/shelf) from an otherwise-intact vault.
N/A Not Yet
Atlas is the memory; you bring the AI. You keep your own subscription and never pay for a second API.
Everything is local. Encrypted sync across your own devices is on the roadmap.
Close the app and your memory is sealed and offline — a feature, not a bug.
Questions, rough edges, "this is annoying because…" — feedback@atlasbrain.io.