Keboola MCP Server: AI Agents, Transformations, SQL

Overview

Keboola MCP Server is an open-source bridge that connects Keboola projects to modern AI assistants and MCP-compatible clients (Cursor, Claude, Windsurf, LangChain adapters, etc.). It exposes Keboola capabilities—storage access, SQL transformations, component operations, jobs, flows and metadata—as callable tools for agents. That lets agents fetch the exact data, run transformations, or trigger jobs without custom glue code.

The server is available as a remote hosted endpoint on Keboola stacks (recommended for most users) and can also be run locally for development. Important: the legacy SSE transport has been deprecated and will be removed on 2026-03-31 — use the Streamable HTTP transport and the /mcp endpoints instead.

Features

• Storage: query tables, read schemas, and manage table/bucket descriptions
• Components: list, inspect, and create extractor/writer/data app and transformation configurations
• SQL: generate and create SQL transformations from natural language prompts
• Jobs: start components/transformations and monitor execution details
• Flows: create and manage workflow pipelines (Orchestrator & Conditional Flows)
• Data Apps: create, deploy and manage Streamlit-based Data Apps scoped to storage data
• Metadata: search, read, and update object/project metadata and documentation
• Dev Branches: operate safely in development branches using the X-Branch-Id header
• OAuth authentication and first-party hosting on Keboola stacks for easy integration

Installation / Configuration

Two common ways to get started: use the Remote MCP Server (easiest) or run a local instance for development.

Remote (recommended)

1. Open Keboola UI → Project Settings → MCP Server
2. Copy the server URL (example): https://mcp.<YOUR_REGION>.keboola.com/mcp
3. Configure your AI assistant / MCP client with that URL
4. Authenticate via OAuth when prompted and select the project

Claude Code: install the Keboola integration with:

# Replace <URL> with the remote server URL for your region
claude mcp add --transport http keboola https://mcp.<YOUR_REGION>.keboola.com/mcp

Region examples:

Local development

• Clone the repo: git clone https://github.com/keboola/keboola-mcp-server
• Follow the repository README to build and run locally (inspect Docker compose or dev scripts)

Branch scoping (Dev branches)

• To scope requests to a development branch, include header:

X-Branch-Id: <branchId>
Authorization: Bearer <oauth-token>

• Without X-Branch-Id the server operates on the production branch.

Migration note

• SSE transport is deprecated. Use Streamable HTTP + /mcp endpoints before 2026-03-31.

Available Tools / Resources

The MCP Server exposes Keboola features as callable tools for agents. Typical tools include:

• storage.query: run SQL or table queries and return results
• transformations.create: produce a SQL transformation (often created from NL)
• components.list / components.get: inspect extractors/writers/apps
• jobs.run / jobs.get: trigger components or transformations and retrieve logs/status
• flows.create / flows.run: build and execute orchestrator or conditional flows
• dataapps.create / dataapps.deploy: create and manage Streamlit Data Apps
• metadata.search / metadata.update: manage project docs and object metadata

Authentication is handled via OAuth; the hosted remote server supports Keboola OAuth flows. Once a user authenticates, agents receive scoped access to the selected Keboola project.

Use Cases (concrete examples)

1. Generate a SQL transformation from plain English

   • Agent receives a prompt: “Create a transformation that aggregates daily sales by region.”
   • Agent calls the MCP SQL tool (transformations.create) to generate SQL and create a transformation object.
   • Developer triggers the transformation via jobs.run and monitors it (jobs.get) until completion.

2. AI-assisted data analysis

   • Agent queries a table with storage.query to retrieve a sample.
   • Agent runs analysis, drafts a report, and optionally creates a Data App to visualize results (dataapps.create + deploy).

3. Triggering and monitoring ETL jobs

   • Use components.list to find a connector or writer.
   • Run it with jobs.run and poll jobs.get for progress and logs — all from the assistant UI.

4. Safe experimentation in a dev branch

   • Provide X-Branch-Id header when configuring the assistant or when making