Dataraum MCP Server

DataRaum

The understanding layer that grounds an organization's operating model in its own data.

A semantic layer tells BI tools what columns are called. DataRaum learns what they mean — the concepts, relationships, rules, and measures of the organization — and grounds each one in the actual data, with a measured confidence behind it. See the docs for the full picture.

Monorepo layout

packages/
├── engine/          # Python — pipeline, detectors, Temporal activity worker
├── cockpit/         # TypeScript — TanStack Start web UI
├── dataraum-config/ # YAML data — entropy config, LLM prompts, verticals (bind-mounted, never imported)
└── infra/           # docker-compose orchestration

Each package has its own README. Start there if you're working in a specific package.

Status

DataRaum runs as a multi-container platform, isolated per workspace:

• engine (Python) — a Temporal activity worker, no HTTP. Does the durable analysis (add_source, begin_session, operating_model) and writes metadata to the workspace's Postgres schema.
• cockpit (TanStack Start) — the web app you use. Hosts the chat agent, renders the results, and orchestrates the journey by triggering engine workflows via Temporal.

Each workspace runs its own pair of those two containers. In front of them sit the portal (the cockpit image in a second role — login, membership routing, workspace provisioning) and Caddy, which serves the portal on the parent domain and each workspace on its own subdomain.

They share one substrate: Postgres (metadata + cockpit state + catalogs), an S3 object store (the DuckLake data lake + uploads), and Temporal (durable orchestration). No HTTP seam between engine and cockpit — the integration surface is Postgres + Temporal. See the platform architecture.

Quick start

# Set the LLM key
cp packages/infra/.env.example packages/infra/.env
echo "ANTHROPIC_API_KEY=sk-ant-..." >> packages/infra/.env

# Bring up the whole installation (substrate, engine worker + cockpit for the
# default workspace, portal, and the Caddy ingress)
docker compose -f packages/infra/docker-compose.yml up -d --wait

# Engine health = the Temporal worker heartbeat (no HTTP endpoint):
docker compose -f packages/infra/docker-compose.yml run --rm --no-deps \
  --entrypoint temporal temporal-admin-tools \
  worker list --namespace default --address temporal:7233   # → Status: Running

# Sign in at the portal, then open the default workspace from there
open http://dataraum.localhost              # dev@dataraum.dev / dataraum-dev

Caddy routes by hostname: the parent domain serves the portal (login + your workspaces), and each workspace has its own subdomain (http://ws1.dataraum.localhost). localhost:3000 is published for debugging only — the session cookie is scoped to the parent domain, so a browser there is redirected to the portal and a script gets 401.

The thing that bites on a first run: Caddy binds port 80. If something already holds it (macOS ships Apache), up fails at container start and the portal never comes up — set CADDY_HTTP_PORT and a matching DATARAUM_PORTAL_ORIGIN to move it. (*.localhost needs no /etc/hosts entry.)

Compose defines exactly one workspace pair — bootstrap scaffolding, so a fresh install has something to log into and something for the provisioner to clone. Every other workspace is created from the portal (New workspace) or bun run workspace:create; compose does not grow a service per workspace.

Full walkthrough, including troubleshooting: Running the stack.

For UI iteration, run the cockpit dev server outside docker for hot reload — see packages/cockpit/README.md.

Run a released version (published images)

The quick start above builds the engine and cockpit from source. To run the published release images instead — a deploy host, no build toolchain — layer the release overlay and name the version:

export DATARAUM_VERSION=1.2.3          # any tag from a GitHub Release
docker compose \
  -f packages/infra/docker-compose.yml -f packages/infra/docker-compose.release.yml \
  --env-file packages/infra/.env up -d --wait --no-build

This pulls ghcr.io/dataraum/{dataraum, dataraum-cockpit, dataraum-cockpit-migrate} at that tag. See Deployment for the images, schema/migration handling, and the per-workspace topology.

Develop

• Engine (Python): cd packages/engine && uv sync --group dev && uv run pytest --testmon tests/unit -q. See packages/engine/README.md and packages/engine/CLAUDE.md.
• Cockpit (TypeScript): cd packages/cockpit && bun install && bun --bun run dev (the --bun flag is required). See packages/cockpit/README.md and packages/cockpit/CLAUDE.md.
• Pull the engine metadata schema (cockpit): cd packages/cockpit && DATARAUM_WORKSPACE_ID=<id> METADATA_DATABASE_URL=<url> bun run db:pull:metadata. Re-run after the engine adds/changes SQLAlchemy models.

Documentation

Platform docs live in docs/ (workspace root) and are published via Zensical. Start at docs/index.md, or serve the site locally:

uv run --project packages/engine zensical serve   # run from the repo root

• Documentation home — what DataRaum does, at a glance
• The journey · pipeline & phases · learnable surface · measurement & detectors
• Platform architecture — under the hood
• Decision records — the why behind the architecture

License

Apache 2.0 — see LICENSE.