Atlas MCP

What is Atlas?

Ask two tools what revenue was last quarter and you can get two different numbers. Atlas reads your definitions first.

What grounds an answer

Three context surfaces, each with a different job — and a boundary between them that Atlas enforces rather than trusts:

That descriptive-vs-authoritative split is the point: a runbook can inform an answer, but only the semantic layer can authorize the SQL behind it.

Every YAML field exists because an LLM needs it to write correct SQL: sample_values ground the agent in real data, glossary.status: ambiguous forces clarifying questions, metrics.objective picks MAX vs MIN, query_patterns teach the canonical join shapes for your domain.

Where the answer shows up

The same grounded agent, reachable however your team already works:

• Chat UI — built in, with the SQL on display behind every answer

• Dashboards — draft-first and publish-gated, so private work stays private until you ship it (ADR-0029)

• MCP server — Claude Desktop, Cursor, Continue, or any MCP client, over stdio or OAuth 2.1

• Embeddable widget — script tag or React component, in your own app

• Chat platforms — six, Slack one-click; Teams, Discord, Telegram, and WhatsApp with your own bot (Google Chat coming soon)

• REST API + CLI — headless, typed, scriptable

Built with Hono, Vercel AI SDK, and bun. Supports Anthropic, OpenAI, Bedrock, Ollama, and Vercel AI Gateway. Works with PostgreSQL, MySQL, ClickHouse, Snowflake, DuckDB, BigQuery, Elasticsearch, and Salesforce.

Related MCP server: Foggy Data MCP Bridge

Try the demo locally

bun create atlas-agent my-app
cd my-app
# edit .env and set your LLM API key (Anthropic / OpenAI / etc.)
bun run dev
# Open http://localhost:3000

The scaffold seeds the canonical NovaMart e-commerce dataset (52 tables, ~480K rows) by default — twelve generic e-commerce KPIs ship as starter prompts inside the chat UI; the canonical 5 below drive the eval harness (#2025) and the docs/landing copy. The scaffold defaults to SQLite + Anthropic; pass --defaults for non-interactive setup or follow the prompts to pick PostgreSQL / OpenAI / etc.

Ask one of the canonical questions in the chat UI:

• "What's our GMV this quarter?"

• "What's our top-performing category by GMV this month?"

• "Monthly GMV trend over the past 6 months."

• "Show me revenue last quarter." — Atlas asks which definition you mean (GMV vs. net revenue vs. seller revenue) because revenue is status: ambiguous in the glossary

• "What are our most common return reasons?"

The agent reads your YAML semantic layer first, picks the right entities, writes SQL, runs it through the validation pipeline, and returns answers with the underlying SQL on display.

The default landing for fresh installs is chat-first — admins can flip to admin in Settings → Profile. See the Default Landing guide for the underlying preference.

Install Atlas as an MCP server

Once you have an Atlas instance (local from the demo above, self-hosted, or a hosted workspace), add it to Claude Desktop, Cursor, or Continue with one command. Auto-detects the client and merges into its config:

bunx @useatlas/mcp init --local            # paste-ready config for a local Atlas instance
bunx @useatlas/mcp init --local --write    # merge into the detected client config (with a .bak)
bunx @useatlas/mcp init --hosted --write   # for an app.useatlas.dev workspace via OAuth 2.1

On WSL2? Bun's bunx shim has resolution issues on some WSL2 setups — substitute bun x (space-separated) for any bunx command above (e.g. bun x @useatlas/mcp init --local). The space-separated form is a bun subcommand and resolves correctly.

Restart Claude Desktop / Cursor and ask the same canonical questions through your AI client. See the MCP guide for the full flow — hosted (mcp.useatlas.dev over OAuth 2.1 + DCR + PKCE) and self-hosted (stdio) live in the same page under tabs.

What's in the YAML?

A 20-line slice of semantic/entities/orders.yml from the bundled NovaMart e-commerce demo (#2021):

name: Orders
type: fact_table
table: orders
grain: one row per order
description: |
  Customer orders — the primary fact table for revenue analysis.
  shipping_cost uses MIXED UNITS (some rows in dollars, some in cents).
dimensions:
  - name: status
    sql: status
    type: string
    sample_values: [pending, processing, shipped, delivered, cancelled]
  - name: order_month
    sql: TO_CHAR(created_at, 'YYYY-MM')
    type: string
    virtual: true
measures:
  - name: total_gmv_cents
    sql: total_cents
    type: sum
joins:
  - target_entity: Customers
    relationship: many_to_one
    join_columns: { from: customer_id, to: id }

That YAML is the contract between your team and the agent — version-controlled, code-reviewed, diffable. Sibling files (glossary.yml, metrics/*.yml, catalog.yml) round it out: glossary terms with status: ambiguous force the agent to clarify, metrics with objective: maximize / minimize make optimization direction explicit, and the catalog routes the agent to the right entity for a given question.

See the full Semantic Layer reference for the complete schema.

Embed in your app

Atlas also ships an embeddable chat widget for any frontend:

<script
  src="https://your-atlas.example.com/widget.js"
  data-api-url="https://your-atlas.example.com"
  data-theme="dark"
></script>

Or use the React component:

import { AtlasChat } from "@useatlas/react";

export default function App() {
  return <AtlasChat apiUrl="https://your-atlas.example.com" />;
}

The widget supports programmatic control (Atlas.open(), Atlas.ask("..."), Atlas.destroy()), event callbacks, and theming. See the widget docs.

Why Atlas?

Deploy

Docker:

git clone https://github.com/AtlasDevHQ/atlas-starter-docker.git && cd atlas-starter-docker
cp .env.example .env   # Set your API key + database URL
docker compose up

How It Works

1. User (or agent) asks a natural language question — over MCP, the chat widget, the API, or a chat platform (Slack, Teams, Discord, Telegram, or WhatsApp; Google Chat coming soon)

2. Agent explores the YAML semantic layer — entities, glossary, metrics, query patterns

3. Agent writes SQL, validated through a 7-layer security pipeline (empty check, regex guard, AST parse, table whitelist, RLS injection, auto-LIMIT, statement timeout)

4. Results are returned with charts and an interpreted narrative

Question → YAML semantic layer → SQL generation → 7-layer validation → Query execution → Charts + narrative

Generate the semantic layer

bun run atlas -- init                 # Profile DB and generate YAMLs
bun run atlas -- init --enrich        # Profile + LLM enrichment
bun run atlas -- init --demo          # Load NovaMart demo data + profile

Architecture

atlas/
├── packages/
│   ├── api/              # @atlas/api — Hono API server + agent loop + tools + auth
│   ├── web/              # @atlas/web — Next.js frontend + chat UI components
│   ├── cli/              # @atlas/cli — CLI (profiler, schema diff, enrichment)
│   ├── mcp/              # @atlas/mcp — MCP server (Claude Desktop, Cursor, etc.)
│   ├── sandbox-sidecar/  # @atlas/sandbox-sidecar — Isolated explore sidecar
│   ├── sdk/              # @useatlas/sdk — TypeScript SDK
│   ├── react/            # @useatlas/react — Embeddable chat component + hooks
│   ├── types/            # @useatlas/types — Shared wire-format types
│   ├── schemas/          # @useatlas/schemas — Shared Zod schemas
│   ├── plugin-sdk/       # @useatlas/plugin-sdk — Plugin type definitions
│   ├── webhook-publisher/# @useatlas/webhook-publisher — HMAC-signed outbound webhooks
│   ├── oauth-helper/     # @atlas/oauth-helper — OAuth 2.1 + DCR + PKCE primitives (internal)
│   ├── okf-bundle/       # @atlas/okf-bundle — OKF knowledge-bundle builder (internal)
│   └── fumadocs-okf/     # @atlas/fumadocs-okf — Fumadocs → OKF adapter (internal)
├── plugins/              # 24 plugins (datasource, context, interaction, action, sandbox)
├── ee/                   # @atlas/ee — Enterprise features (source-available, commercial license)
├── create-atlas/         # Scaffolding CLI (bun create atlas-agent)
├── apps/
│   ├── www/              # Landing page (useatlas.dev)
│   └── docs/             # Documentation (docs.useatlas.dev)
└── examples/             # Docker + Vercel deploy examples

Security

SQL validation runs through multiple layers. Your database credentials and query results never leave your infrastructure — only questions reach the LLM provider (use Ollama for fully self-hosted).

See sandbox architecture for the full threat model.

Environment Variables

See .env.example for all options.

Documentation

• The Semantic Layer — Entities, dimensions, measures, joins, glossary, metrics — the YAML format reference

• Knowledge Base — Mirror your own docs as descriptive context, via connectors or upload

• Dashboards — Draft-first, publish-gated dashboards built from chat answers

• MCP Server — Use Atlas from Claude Desktop, Cursor, Continue

• Quick Start — Local dev from zero to asking questions

• Demo Dataset — NovaMart e-commerce dataset and canonical questions

• Deploy Options — Docker, Railway, Vercel, and more

• Connect Your Data — Connect to an existing database safely

• Widget Embedding — Script tag and React component

• Bring Your Own Frontend — Nuxt, SvelteKit, React/Vite, TanStack Start

• Plugin Authoring — Build custom plugins

• Security & Sandbox — Threat model, isolation tiers

• Enterprise Boundary — /ee features, AGPL vs commercial split, requireEnterprise API

Contributing

Quick development setup:

git clone https://github.com/AtlasDevHQ/atlas.git && cd atlas
bun install
bun run db:up         # Start Postgres + sandbox sidecar
cp .env.example .env  # Set ATLAS_PROVIDER + API key
bun run dev           # http://localhost:3000

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 Atlas server and core packages (@atlas/api, @atlas/cli, @atlas/web, @atlas/mcp, @atlas/sandbox-sidecar) are licensed under AGPL-3.0. If you modify the server and serve it to users, you must share those modifications.

The client libraries (@useatlas/sdk, @useatlas/react, @useatlas/types, @useatlas/plugin-sdk) and all plugins are licensed under MIT. Embed them in proprietary apps with no restrictions.

The ee/ directory (@atlas/ee — SSO, SCIM, custom roles, approval workflows, residency, branding, and the rest of the SaaS surfaces) is source-available under a commercial license. Self-hosted users get the full AGPL core for free; the commercial license adds enterprise governance and the polished hosted experience. See the Enterprise Boundary page for the full feature inventory.