Keboola MCP Server – AI Data Pipeline Control

Keboola MCP Server connects MCP-compatible assistants and agents to a Keboola project. It exposes project data operations, component configuration, SQL transformation, job execution, flows, data apps, metadata, and documentation tools while respecting the selected project, branch, credentials, and tool restrictions.

#data-pipelines#sql#orchestration

Overview

Keboola MCP Server is the official Model Context Protocol server for the Keboola data platform. It acts as a bridge between MCP clients such as Cursor, Claude, Windsurf, VS Code, and custom AI agents and a selected Keboola project. The server is intended for teams that want an AI assistant to work directly with governed data infrastructure instead of relying only on pasted context or static documentation.

What the MCP server enables

The server exposes Keboola capabilities as MCP tools. Verified capabilities include querying Storage tables, managing bucket and table descriptions, listing and inspecting component configurations, creating SQL transformations, running jobs, retrieving job execution details, building workflow pipelines with Conditional Flows and Orchestrator Flows, managing Streamlit Data Apps, searching or updating project metadata, and using development branches for safer changes outside production. The server supports tools and prompts, but Keboola documentation states that formal MCP Resources, sampling, and roots are not currently supported server features.

When to use it

Use this MCP server when an AI agent needs to help build, debug, document, or operate Keboola data pipelines. Practical examples include asking an assistant to inspect project storage, generate a SQL transformation, run or troubleshoot a component job, create a data app over warehouse data, or find official Keboola documentation while working in an AI client. Development branches are useful when the agent should propose or implement changes without immediately affecting production.

Connection and authentication

Keboola supports Streamable HTTP and stdio. Streamable HTTP is the recommended remote option and is hosted on every multi-tenant Keboola stack, with URLs such as https://mcp.keboola.com/mcp or region-specific https://mcp.<YOUR_REGION>.keboola.com/mcp; OAuth is used by the remote server so the connected tools match the authenticated user's Keboola role. Local stdio execution can be run with uvx keboola_mcp_server --api-url $KBC_API_URL and requires KBC_STORAGE_TOKEN, KBC_WORKSPACE_SCHEMA, and KBC_API_URL. BigQuery-backed projects may also require GOOGLE_APPLICATION_CREDENTIALS for local execution.

Key considerations

SSE is documented as deprecated and scheduled for removal on 01.04.2026, so it is intentionally not listed as a supported transport here. HTTP deployments can restrict tool availability with X-Allowed-Tools, X-Disallowed-Tools, and X-Read-Only-Mode, but those authorization headers apply only to HTTP transports, not local stdio. For highly controlled permissions, Keboola recommends local deployment with a specific Storage API token and workspace schema. Treat the storage token and warehouse credentials as sensitive secrets and scope them to the minimum project access required.

Frequently Asked Questions

When should an AI agent use the Keboola MCP Server?

Use it when an AI agent needs to inspect, build, document, or operate data workflows inside a Keboola project. It is suited for tasks such as querying Storage tables, creating SQL transformations, running jobs, debugging executions, or working in development branches before production changes.

What does this MCP server add to an AI agent's capabilities?

It gives the agent live Keboola tools for storage operations, component management, SQL execution, job control, flows, Data Apps, metadata, documentation search, and project information. This is different from model knowledge because the tools operate against the connected Keboola project and its current state.

What can the AI agent access or manage through this server?

Depending on credentials and allowed tools, the agent can query tables, inspect buckets, list or create component configurations, run components and transformations, retrieve job details, manage workflow pipelines, create Streamlit Data Apps, and search or update project documentation and metadata. Formal MCP Resources are not currently exposed by the server.

How is authentication configured for this MCP server?

Remote Streamable HTTP uses Keboola OAuth and prompts the user to log in, select a project, and authorize access. Local stdio uses a Keboola Storage API token, workspace schema, and Keboola API URL through `KBC_STORAGE_TOKEN`, `KBC_WORKSPACE_SCHEMA`, and `KBC_API_URL`; BigQuery projects may also need Google application credentials.

Which transport should be used for this MCP server?

Streamable HTTP is the recommended remote option for clients that support remote OAuth. Local stdio is useful for development, testing, or permission control with a specific Storage API token and workspace schema. SSE is not included because Keboola documents it as deprecated and scheduled for removal on 01.04.2026.