mcpField
A
Atlas

Atlas

Local
docs.useatlas.devData & Analytics1Updated 6w ago
RepositoryVisit website

About Atlas

Atlas is an open-source semantic layer for company analytics, designed for AI agents and self-hosting on a machine or in a VPC. It distinguishes live data queries, human-approved facts, and raw source material, while preserving provenance for answers.

Key features

  1. 01YAML-authored analytics semantic layer
  2. 02Answers distinguish three evidence types
  3. 03Live, validated read-only data queries
  4. 04Human approval for attested facts
  5. 05Contradictions retain both claims and sources

Use cases

  1. 01Ask an agent questions about company data
  2. 02Trace a fact to its source and approver
  3. 03Inspect conflicting claims without choosing a winner

MCP support

Atlas provides an MCP server that AI clients can connect to locally or to a self-hosted instance. The README gives an initialization command for configuring clients such as Claude Desktop, Cursor, and Continue.

Configuration

Local server config is the same for every client. Copy it into your client’s MCP config file.

mcp.json · Universal config
{
  "mcpServers": {
    "atlasdevhq-atlas": {
      "command": "npx",
      "args": [
        "-y",
        "@useatlas/mcp"
      ],
      "env": {
        "ATLAS_DATASOURCE_URL": "<ATLAS_DATASOURCE_URL>",
        "ATLAS_PROVIDER": "<ATLAS_PROVIDER>",
        "ATLAS_API_URL": "<ATLAS_API_URL>"
      }
    }
  }
}
Run commandnpx -y @useatlas/mcp
Environment variables to set
ATLAS_DATASOURCE_URLSecretAnalytics datasource URL (postgres, mysql, clickhouse, snowflake, duckdb, bigquery, salesforce). Optional — falls back to a bundled NovaMart SQLite demo when unset.
ATLAS_PROVIDERLLM provider: anthropic, openai, bedrock, ollama, openai-compatible, gateway. Optional — the MCP server does not call an LLM directly (the client does); set this only if you also run Atlas's chat or scheduler.
ATLAS_API_URLURL of a running Atlas API. Optional — defaults to http://localhost:3001. When set, the MCP server inherits the API's connections, semantic layer, and governance.

Local servers need a runtime on your machine (e.g. Node.js 18+). Config file locations differ by client; see your client’s docs.

Packages
npm@useatlas/mcp0.0.2

Tools

No tool list yet

Local servers expose tools only once running on your machine, and the maintainer has not provided a tool list.

Details

From the project README

Atlas

Atlas is the company facts your AI agents can trust: every one carries its source, its date, and the name of the person who approved it. Open source, runs in your VPC.

CI npm License

The hosted service is shut down. I ran Atlas as a hosted service at useatlas.dev until September 2026 and have now turned it off, along with the hosted demo and docs.useatlas.dev. Atlas continues as open source: everything below runs on your own machine or in your own VPC. The docs live in apps/docs/content.

Point Claude Desktop, Cursor or Continue at a local Atlas with one command. With no datasource configured it uses the bundled NovaMart demo data:

bunx @useatlas/mcp init --local --write

Restart the client and ask it a question about NovaMart's orders. (Claude Code, VS Code and other clients: the same command with --client generic prints a block to paste. On WSL2, write bun x instead of bunx.)

How it works

Three kinds of thing live in the Atlas, and every answer says which it is drawing on:

  1. Surveyed — read straight from your company's own data through a semantic layer you author. True by construction: the query re-reads live rows, nobody interpreted anything, and it cannot go stale between readings. SELECT-only, single statement, table-whitelisted, validated seven ways before it runs.
  2. Attested — extracted from something someone wrote, then approved by a named person in your company. That person is on the record, and the fact carries its source and its date.
  3. On the record — the raw source material itself, unedited. Trustworthy as testimony, not as fact.

Surveyed outranks Attested wherever they overlap, so a recollection never overwrites the data. Nothing becomes Attested without a person approving it, and there is no setting that turns that off. Contradictions are shown with both claims and both sources; Atlas does not pick a winner. Where nobody has surveyed, the coverage page says so instead of guessing.

Run it

  • Self-hosted — the complete Atlas under AGPL, in your VPC, free; Docker, Railway or Vercel. Self-host quick start.
  • From your AI agent — bunx @useatlas/mcp init --local against a local Atlas, or bunx @useatlas/mcp init --hosted --api-url <your Atlas> against a self-hosted one with managed auth. MCP guide.

Where everything else went

This README used to inventory the whole product. Each section now lives on one docs page, in the docs source:

Acknowledgments

Atlas was inspired by Abhi Sivasailam's work on Vercel's internal data agent d0 and the open-source vercel-labs/oss-data-analyst template. The core insight — invest in a rich semantic layer, trust the model, and keep the tool surface minimal — came from that work.

License

The server and core packages are AGPL-3.0: if you modify the server and serve it to users, you share those modifications. The client libraries (@useatlas/sdk, @useatlas/react, @useatlas/types, @useatlas/plugin-sdk, @useatlas/mcp) and all plugins are MIT. The ee/ directory is source-available under a commercial license; nothing that makes the Atlas work is behind it — only governance, convenience and scale. Full inventory: Enterprise boundary.

Comments

Markdown supported

New to MCP?

From what the protocol is to installing your first server, these answers cover most first questions.

An open protocol that defines how AI applications connect to external tools and data in a uniform way. Clients talk to the model; MCP servers provide tools, resources and prompt templates. They exchange standard messages, so one server can be reused by many clients.