Claude-Atlas-MCP

Self-hosted MCP server that gives Claude persistent memory across conversations — entities, observations, history, and time-based reminders, in a lightweight Node/SQLite backend you run yourself.

Point Claude at it as an MCP connector and it can remember what you're working on from one conversation to the next: ongoing projects, decisions and their rationale, facts about you and your setup, and things to resurface on a future date.

Why

Claude forgets everything when a conversation ends. Atlas is a small, boring, durable memory layer you own end to end — no third-party service, no vendor lock-in. It's a single Node process backed by one SQLite file. Run it on a home server, a VPS, or your laptop.

Related MCP server: claude-memory-mcp

Data model

Tools

The server exposes 16 MCP tools:

Reading

• get_landscape — everything in a section (with shared auto-merged in): all entities with their observations, plus any due reminders. Call at the start of a conversation to get oriented.

• search — keyword search across entities, observations, and history.

• get_entity — one entity and its observations by name.

• get_history — the timeline of logged events.

Writing

• upsert_entity — create or update an entity's name/summary.

• add_observation — attach a fact to an entity.

• update_observation — edit a fact in place; the ID and its history stay stable. Works on protected rows.

• remove_observation — drop a fact that's stale or done (refuses if protected).

• protect_observation / unprotect_observation — mark a fact as protected from deletion, or lift that protection.

• remove_entity — delete an entity and its observations (refuses if any observation is protected).

• log_event — record a notable event to history.

Reminders

• create_reminder — a note with a trigger_date (and optional entity link).

• list_reminders — all reminders, or just those currently due.

• dismiss_reminder — mark a reminder handled (it stops surfacing).

• remove_reminder — delete a reminder outright.

Every tool response also carries a small time footer — current server time (America/New_York) plus elapsed time since that token's last call — so the model never has to guess or do date math from a stale mental clock.

Groom worker

src/groom.js is a standalone script meant to run on a schedule (cron, or docker exec <container> node src/groom.js). It's intentionally report-only and mechanical — no LLM calls, no deletion of your data:

• flags likely near-duplicate observations within an entity

• flags dormant entities (60+ days untouched) as archive/compress candidates

• flags long-dismissed reminders (90+ days) as removal candidates

• rotates its own audit_log (90+ days) — the only thing it actually deletes

• skips entities untouched since the last run, so repeat runs are cheap

Findings land in a per-section "Groom Report" entity for you (or Claude) to act on.

Requirements

• Node.js with built-in SQLite support (node:sqlite) — Node 22+.

• Docker (optional, recommended for deployment).

There are no native dependencies — storage is the built-in node:sqlite module, so there's nothing to compile.

Quick start (Docker)

git clone https://github.com/dcazman/Claude-Atlas-MCP.git
cd Claude-Atlas-MCP
cp .env.example .env      # set a token
docker compose up -d --build

The server listens on port 7784. Data persists to ./data (a single SQLite file) via the mounted volume.

Quick start (bare Node)

Requires Node 22+ (for built-in node:sqlite).

npm install
cp .env.example .env      # set a token
npm start                 # or: node src/server.js

Connecting Claude

Atlas speaks MCP over streamable HTTP at POST /atlas-mcp. Add it as a connector using the server's URL with your token:

https://<your-host>/atlas-mcp?token=<your-secret>

The token is the secret half of an ATLAS_TOKEN caller:secret pair (see below). You can also pass it as an X-Atlas-Token header or a Bearer token instead of the query string.

There's no section in the URL — every tool takes a section argument (work, personal, or shared), and which one a given conversation should default to is best set in your Claude project's custom instructions (e.g. "Your Atlas section is personal"). A GET /health endpoint is available for liveness checks.

For real use you'll want it behind HTTPS — a reverse proxy or a tunnel (Cloudflare Tunnel, Tailscale, nginx, etc.) in front of the container. The token is the only auth, so do not expose the port publicly without TLS.

Once connected, a good habit is to have Claude call get_landscape at the start of each conversation and keep entries updated (upsert_entity / add_observation / log_event / remove_observation) as things change — you can encode that in the connector's instructions field or your project's custom instructions.

Securing it

Atlas's built-in auth is a single shared token — fine behind a private network or tunnel, but thin if you're exposing it to the internet. For real access control, put a dedicated auth gateway in front rather than hardening this server yourself.

mcp-auth-proxy is a drop-in OAuth 2.1 / OIDC gateway for MCP servers — no code changes to Atlas:

• Authenticate against your own IdP (Google, GitHub, Okta, Auth0, Azure AD, Keycloak, any OIDC provider), with an optional password.

• Authorize users by exact match or glob (e.g. *@yourcompany.com).

• Terminates TLS and proxies HTTP transports through as-is, verified across Claude, Claude Code, ChatGPT, Copilot, and Cursor.

Roughly, you'd point it at Atlas's HTTP endpoint:

./mcp-auth-proxy \
  --external-url https://<your-domain> \
  --tls-accept-tos \
  -- http://localhost:7784/atlas-mcp

See its documentation for IdP setup and configuration. (Not affiliated — just a clean fit for self-hosted MCP servers like this one.)

Configuration

Set via .env (see .env.example):

Security

See SECURITY.md for the threat model, deployment hardening notes, and how to report a vulnerability.

License

MIT — see LICENSE.