databricks-transform-mcp

A Databricks App that hosts the Coalesce Transform MCP server over HTTP so Databricks Genie can connect to it.

How it works

Genie / AI Agent
      │
      │  HTTPS  (Databricks App proxy handles auth)
      ▼
Databricks App  (uvicorn + server.py)
      │
      │  Python MCP SDK bridges HTTP ↔ stdio
      ▼
coalesce-transform-mcp  (npm package, 100+ tools)
      │
      ▼
Coalesce API  +  Snowflake

server.py uses the Python MCP SDK to spin up coalesce-transform-mcp as a subprocess, then exposes its tools over Streamable HTTP via a Starlette/uvicorn ASGI app.

Endpoint	Path	Use for
Streamable HTTP	/mcp	Genie, newer MCP clients
Health	/	Databricks App health checks

---

Prerequisites

• Databricks workspace with Apps enabled
• Coalesce access token (Coalesce Deploy → Settings → API Tokens)
• Node.js ≥ 22 on the App cluster (needed to run npx coalesce-transform-mcp)
• Python ≥ 3.11

---

1. Store the Coalesce token in Databricks Secrets

# Create a secret scope (skip if it already exists)
databricks secrets create-scope coalesce-mcp

# Store your Coalesce API token
databricks secrets put-secret coalesce-mcp coalesce-access-token
# paste your token when prompted

---

2. Link the secret to the app via the UI

Databricks Apps injects secrets as env vars only after they're linked to the app through the UI:

1. In Databricks, go to Compute → Apps and open (or create) the app
2. Click Edit app → App resources → + Add resource → Secret
3. Set:
   • Scope: coalesce-mcp
   • Key: coalesce-access-token
   • Resource key: coalesce-access-token ← must match valueFrom in app.yaml
   • Permission: Can read
4. Save and redeploy

This is what wires the COALESCE_ACCESS_TOKEN env var in app.yaml to the actual secret value.

---

3. Deploy

Push this repo to a Databricks Git Folder, then deploy:

databricks apps deploy coalesce-transform-mcp \
  --source-code-path /Workspace/Repos/<your-user>/databricks-transform-mcp

Or via the Apps UI:

1. Compute → Apps → Create App → Custom
2. Point to this Git Folder
3. Click Deploy

On first deploy, pip install -r requirements.txt runs automatically, then uvicorn starts.

---

3. Connect Genie

1. In Genie, go to Settings → MCP Servers → Add
2. Set the endpoint to the Streamable HTTP path:

   https://<workspace-host>/apps/coalesce-transform-mcp/mcp
   

3. Auth is handled automatically by the Databricks App proxy

---

Environment Variables

Variable	Required	Description
COALESCE_ACCESS_TOKEN	Yes	Bearer token from Coalesce Deploy
COALESCE_BASE_URL	No	Region base URL (default: https://app.coalescesoftware.io)
COALESCE_REPO_PATH	No	Path to local Coalesce repo for repo-backed tools
COALESCE_MCP_READ_ONLY	No	Set "true" to disable write/delete tools
SNOWFLAKE_ACCOUNT	No	Snowflake account for run/create operations
SNOWFLAKE_USER	No	Snowflake user
SNOWFLAKE_PRIVATE_KEY	No	Snowflake private key (PEM)

---

Local Testing

pip install -r requirements.txt

export COALESCE_ACCESS_TOKEN=your-token
export COALESCE_BASE_URL=https://app.australia-southeast1.gcp.coalescesoftware.io

uvicorn server:app --host 0.0.0.0 --port 8000

Test the MCP endpoint:

curl -X POST http://localhost:8000/mcp \
  -H "Content-Type: application/json" \
  -d '{"jsonrpc":"2.0","method":"tools/list","id":1}'