ATLAS MCP Server
The ATLAS MCP server keeps an agent's projects, tasks and knowledge in a Neo4j graph database: dependencies, search and a deep research plan
- 478 ★ on the repository
- Last commit
- Apache-2.0
Broad access
Why this level
- Creates, changes and deletes projects, tasks and knowledge in your Neo4j database
- The atlas_database_clean tool wipes the whole database in one call
- Keeps the Neo4j password in the client config or in .env
Install
- Works in
Manual install
git clone https://github.com/cyanheads/atlas-mcp-server.git && cd atlas-mcp-server && npm install && docker-compose up -d && npm run buildSet the Neo4j parameters in .env before running. The server builds to dist/index.js.
Access and checks
How we reviewBroad accessThe tool can change data, run commands or handle secrets.
- Creates, changes and deletes projects, tasks and knowledge in your Neo4j database
- The atlas_database_clean tool wipes the whole database in one call
- Keeps the Neo4j password in the client config or in .env
This is third-party code. Review the repository files before installing.
What it does
ATLAS (Adaptive Task and Logic Automation System) gives an agent project, task and knowledge management in three linked tiers. A project contains tasks, and knowledge items with tags, a domain and citations attach to projects and tasks. Projects and tasks can depend on each other, and the server validates those links. Data lives in Neo4j. The tools are atlas_project_create, list, update and delete, atlas_task_create, update, delete and list, and atlas_knowledge_add, delete and list. Create, update and delete work one at a time or in bulk. The cross-entity atlas_unified_search searches all three entities with full-text or a regular expression on a chosen property. atlas_deep_research creates a hierarchical research plan in the database for a topic and goal. The same data is exposed as atlas://projects, atlas://tasks and atlas://knowledge resources. The server can back up and restore the database, keeping up to 10 copies in ./atlas-backups by default. The default transport is stdio, and HTTP on port 3010 is enabled with an environment variable. An experimental web UI is included.
Who it is for. For developers running long multi-step projects with an agent who want tasks, dependencies and findings kept in a graph database instead of in prompt text.
Good fit when
- A project has dozens of tasks with dependencies, and the agent must see what depends on what and in which state
- Research findings should be stored as knowledge items with tags, a domain and source citations
- You want one search over projects, tasks and knowledge, including a regex on a chosen field
- A deep research needs a structured plan in the database first: topic, goal, scope
- Data needs backups and a way to restore them
Not a fit when
- You cannot run Neo4j: the server does not work without a database
- You need a simple to-do list in one file: a database and MCP are overkill
- You plan to expose the server as a production service on a network: the developer notes production use has not been tested
- The agent must not be trusted with atlas_database_clean, which wipes the whole database, and you cannot disable it
Example request
Create an ATLAS project for a billing migration, add five tasks with dependencies and store the constraints you find as knowledge itemsMCP
- Transport
- stdio, http
- Authentication
- not required
Environment variables
NEO4J_URIrequired- Neo4j connection address, for example bolt://localhost:7687
NEO4J_USERrequired- Neo4j user
NEO4J_PASSWORDrequired, secret- Neo4j password
MCP_TRANSPORT_TYPE- stdio (default) or http
MCP_AUTH_SECRET_KEYsecret- A secret of at least 32 characters for JWT on the HTTP transport
Limitations
- A running Neo4j instance is required; the repository ships a docker-compose.yml to start one.
- Connection settings NEO4J_URI, NEO4J_USER and NEO4J_PASSWORD go into the client config or .env.
- The HTTP transport supports JWT protection via MCP_AUTH_SECRET_KEY, but the developer notes production use has not been tested.
- The latest changes in the repository date from July 2025.
- Restoring from a backup first clears the current database completely, so it overwrites all data.
How to disable. Remove the atlas-mcp-server block from the MCP client config and stop the Neo4j container with docker-compose down. Data stays in the Neo4j volume and in the ./atlas-backups folder, which you can delete by hand.
FAQ
How do I connect ATLAS MCP to Claude Desktop?
Build the project, start Neo4j and add a block to claude_desktop_config.json with command node, args the path to dist/index.js and env with NEO4J_URI, NEO4J_USER, NEO4J_PASSWORD and MCP_TRANSPORT_TYPE=stdio. The path to dist/index.js must be absolute.
Why does ATLAS need Neo4j?
Projects, tasks and knowledge are linked by relationships and dependencies, and Neo4j stores them as a graph with ACID transactions. That lets the server validate dependencies and search properties with fuzzy matching and wildcards.
What does the atlas_deep_research tool do?
It creates a hierarchical research plan in the database for a project from a topic, a goal and optional scope. The plan lays the work out as tasks and knowledge items the agent then works through.
What does atlas_database_clean do?
It fully resets the database and removes all projects, tasks and knowledge. The tool is marked destructive in its description and requires an acknowledgement parameter set to true. Make a backup first.
How do I back up the ATLAS database?
Run npm run db:backup: it exports projects, tasks, knowledge and relationships to JSON files and a combined full-export.json in a timestamped folder under atlas-backups. npm run db:import with the backup folder path restores it, but that overwrites all current data.
Discussion
No questions or reviews yet.
Related
MCP serversMCP server for the knowledge base Beever Atlas builds from Slack, Discord, Teams and Mattermost chats: cited answers, channel wikis, a graph
An MCP server for Jira and Confluence, cloud and self-hosted: search, create and edit issues and pages
Official Anytype MCP server: an agent searches, creates and edits objects, lists and types in your knowledge base
Jira MCP in Go: an agent searches issues by JQL, runs sprints, comments and worklogs and sees an issue's branches and pull requests on Cloud and Server
The official Todoist MCP server: the agent finds, creates and updates tasks and projects in your account
An MCP server for Linear: issues, projects, cycles, milestones, documents and project updates via the GraphQL API
Comparison with similar tools
Entries that share the main search query, and how they differ.
| Entry | Official | Authentication | Repository stars | Main difference |
|---|---|---|---|---|
| ATLAS MCP Server (this entry) | no | not required | 478 | |
| MCP Shrimp Task Manager | no | not required | 2,148 | turns a request into a step plan with dependencies, with no graph database or knowledge items |
| Taskmaster | no | not required | 28,277 | breaks a PRD into dependent development tasks, while ATLAS stores projects, tasks and knowledge in Neo4j |
| AgentRQ | yes | API key | 1,170 | a self-hosted task manager for human and agent collaboration |
README in short
The page describes ATLAS as a project, knowledge and task management system for AI agents on three graph nodes, with an ASCII diagram of the links. It then covers integration with MCP clients, a feature table, installation, running, environment variables and configuration examples for stdio and HTTP. The bulk is tables of tools and resources, ending with backup and restore sections and a link to examples.