Databricks MCP Server
Connect to the Databricks MCP server to run SQL, orchestrate jobs, manage clusters, and query models and vector indexes across your lakehouse using AI agents on Gumloop, Claude, or Cursor.
Try Databricks now
Type what you want done. Sign in and run it live with an AI agent.
|
Databricks
Installation
Get StartedCreate a Gumloop Account
To use this MCP, you need a Gumloop account. If you don't have one yet, you can create one for free.
Copy Your Server URL
Copy your MCP server URL and add it to your client. You'll be prompted to authorize on first use.
Tools (14)
Get Me
Get information about the authenticated user. Returns the current user's details including username, email, and account information.
List Clusters
List all pinned and active clusters, and all clusters terminated within the last 30 days.
Terminate Cluster
Terminates a Spark cluster with the specified ID. Cluster is removed asynchronously and will be in TERMINATED state when complete.
Start Cluster
Starts a terminated Spark cluster. Preserves previous cluster ID and attributes. Starts with last specified size or minimum nodes if autoscaling.
List Jobs
Retrieves a list of jobs in the workspace.
Manage Job Run
Cancel a running job or delete a non-active job run.
Run Job
Trigger a new job run and return the run_id. Supports idempotency tokens to prevent duplicate runs.
Get Job Run Output
Retrieve the output and metadata of a single task run. Returns first 5 MB of output. For notebook tasks, gets the value from dbutils.notebook.exit().
Query Serving Endpoint
Query a serving endpoint with input data. Supports various input formats including dataframes, tensors, and prompts for external/foundation models.
Query Vector Index
Query a vector index for similarity search. Supports ANN and HYBRID search with optional filtering.
Execute Sql
Execute a SQL statement and optionally await its results. Supports inline results (<25 MiB) or external links (up to 100 GiB).
List Warehouses
List all SQL warehouses that you have access to.
What is Databricks MCP?
The Databricks MCP server gives AI agents access to your Databricks workspace. That means agents can list, start, and terminate Spark clusters, trigger and cancel job runs, pull job output, run arbitrary SQL against your warehouses, query deployed model and foundation LLM endpoints, and run similarity search over your vector indexes. It works across the core objects of your lakehouse: clusters, jobs, SQL warehouses, serving endpoints, and vector search indexes.
If your data engineers, data scientists, and ML teams spend time switching between the Databricks UI and the REST API to kick off pipelines, babysit cluster state, copy query results around, or wire up model calls by hand, an AI agent can take over a lot of that busywork. Describe what you need, and your AI agent will handle the run for you.
MCP stands for Model Context Protocol. It’s an open standard that gives AI agents a way to connect to external tools and services. Instead of coding against the Databricks REST API, handling OAuth token exchange, and parsing statement responses yourself, you connect your Databricks workspace once by adding your workspace URL and OAuth app credentials. After that, you can operate your lakehouse just by chatting with your AI agent.
Related MCP servers
Google BigQuery MCP Server
Query and manage BigQuery datasets, tables, and execute SQL
Snowflake MCP Server
Interact with Snowflake data warehouse for executing SQL queries, managing tables/databases, and accessing data
ClickHouse MCP Server
Query ClickHouse Cloud services via SQL, manage services, backups, ClickPipes and costs, and manage ClickStack observability (dashboards, alerts, sources)
PostgreSQL MCP Server
Query and manage PostgreSQL databases with schema exploration, SQL execution, and query analysis
Looker MCP Server
Interact with Looker Enterprise dashboards and analytics content
Tableau MCP Server
Interact with Tableau to access views, workbooks, datasources, and Pulse metrics
What you can do with Databricks MCP on Gumloop
Manage Spark cluster lifecycle
List your pinned, active, and recently terminated clusters, then start a terminated cluster or terminate a running one by ID. Starting and terminating preserve each cluster’s existing configuration, so an agent can bring compute up before a job and shut it down after.
Trigger jobs idempotently and selectively
Run an existing job by ID with runtime parameter overrides, and pass an idempotency token (up to 64 characters) to prevent duplicate runs in an agent loop. Use the selective task option to trigger only specific task keys within a multi-task job for fine-grained pipeline control.
Cancel runs and pull run output
Cancel an active job run or delete a finished run record, then retrieve the output and metadata of a single task run (first 5 MB), including the value returned by a notebook task. Your agent can monitor a run and react when it finishes or stalls.
Execute arbitrary SQL on your warehouses
List your accessible SQL warehouses and run any SQL statement (up to 16 MiB) against them, covering DDL, DML, and DQL. Choose synchronous or asynchronous execution, set row and byte limits, pick JSON, Arrow, or CSV output, and use external links to handle results up to 100 GiB.
Query Databricks-hosted models and foundation LLMs
List your model serving endpoints and query a deployed model for inference. Send a prompt or chat messages with max tokens and temperature to call Databricks-hosted foundation models like DBRX or Llama, or pass dataframe records for classic predictions and embeddings.
Run hybrid vector search for RAG
List your vector search endpoints and run similarity search against an index using either a raw embedding vector or natural language text. Pick ANN or hybrid (vector plus keyword) query modes, apply JSON filters, and set the number of results, which makes it a clean retrieval step for RAG pipelines.
Check identity and entitlements
Pull the authenticated user’s details, including username, email, groups, and entitlements, so an agent can confirm which account and access it is operating under before it acts.
How to connect the Gumloop Databricks MCP Server
- 1
Create a free Gumloop account
Sign up at gumloop.com. No credit card required.
- 2
Add the Databricks MCP server
Copy your MCP server URL from Gumloop and add it to your preferred client (Claude, Cursor, or Gumloop workflows). You'll authorize on first use.
- 3
Start using Databricks in your AI workflows
That's it. Your AI agent can now run SQL, trigger and manage jobs, control clusters, and query models and vector indexes across your workspace. Use it inside a Gumloop automation, in Claude Desktop, or in Cursor.
Databricks MCP use cases
Pipeline orchestration for data engineers
When upstream data lands, a Gumloop agent can start the right cluster, trigger a Databricks job with an idempotency token so it never double-fires, poll for completion, and pull the run output. If a task fails, the agent posts the error to Slack and reruns only the failed task using the selective task option, so engineers stop babysitting runs.
Self-serve SQL analytics for analytics teams
An analyst describes the numbers they need, and an AI agent writes the SQL, runs it against a SQL warehouse, and returns the results. For large pulls, the agent uses external links and CSV output, then drops a summary into Google Sheets or a Slack message, so the team gets answers without writing queries by hand.
RAG over your lakehouse for ML engineers
Build a retrieval agent that takes a question, runs a hybrid vector search against your Databricks vector index for the most relevant chunks, and passes them to a Databricks-hosted foundation model like DBRX or Llama through a serving endpoint. The agent returns a grounded answer and can log each question and source to a Google Doc for review.
Cost-aware cluster management for platform teams
A scheduled Gumloop agent can list all clusters, terminate ones that are running but idle outside business hours, and post a short cost-hygiene summary to Slack. Before a nightly job window, the same agent starts the clusters that batch pipelines depend on, so compute is up exactly when it’s needed.
Model-backed reporting for data scientists
After a training job completes, an agent can query the new serving endpoint with a held-out set of dataframe records, compare predictions against a SQL query of recent outcomes, and compile a model performance summary. It writes the report to Google Docs and flags drift in Slack, so scientists review results instead of assembling them.
Why use Gumloop for Databricks MCP
Connect your workspace once, no token juggling
Connect your Databricks workspace once by adding your workspace URL and OAuth app credentials. After that Gumloop handles the token exchange and refreshes your tokens automatically. No env vars, token juggling, or coding against the Databricks REST API necessary.
Works with multiple MCP clients
Use the Databricks MCP server endpoint in Claude Desktop, Cursor, or directly inside Gumloop agents. Same server URL, works with any MCP client.
Chain Databricks with 100+ other integrations
Combine Databricks with Slack, Gmail, Google Sheets, Google Docs, and other MCP tools in a single AI agent. An agent can run a query in Databricks, process the results with an LLM, and write them back to your warehouse or out to anywhere else.
Enterprise-grade and scalable
Built for teams, with role-based permissions and dedicated support for Pro users. For details on Gumloop’s security practices, see trust.gumloop.com.
Pricing includes a free plan
You can test the Databricks MCP integration on Gumloop’s free tier before committing. Paid plans start at $37/month.
Frequently asked questions
Ship Databricks agents in minutes
Connect any AI agent to 100+ MCP servers, zero setup.