Introducing Keboola CLI
CLI that speaks AI.
All your projects from one terminal. AI agents are bounded by a permission firewall that can't touch production.
scripted preview · nothing here touches a real project
Interactive demo: you ask Claude Code something in plain English and it drives the Keboola CLI for you. For example, "find the last failed job in prod and fix the SQL": the agent runs kbagent job list and job detail to find the failure, opens a workspace from the transformation to reproduce it, stages the fix with config update --dry-run, and stops there, because config changes and job runs are always separate steps. Its attempt to delete the workspace is blocked by the permission firewall (cli:destructive). Other examples: which projects still read the legacy orders bucket (one search fans out across all 30 connected projects), comparing a transformation between prod and a dev branch, and finding orphaned workspaces.
What is the Keboola CLI
The whole platform, in a command line.
Everything you can click through in the UI: projects, configurations, jobs, storage, transformations, flows, dev branches and workspaces. Plus what the UI cannot do at all: bulk operations across dozens of projects at once, configurations versioned in Git as files, and lineage across projects.
It was built for agents from the start, not a CLI for humans with AI added later. That is why every command returns structured JSON, exit codes and errors, supports a dry run, and never stops to wait at a prompt.
What you get
Search 30 projects at once
Find a table, config, flow or data app across every connected project in one call.
AI agents you can trust
Preview a change before it runs, or lock the session read-only. Enforced in code.
Test SQL code without running it
Load the transformation's real input data into a temporary workspace and iterate there. Query, fix it, then apply.
GitOps for Keboola
Configs as YAML files. Branch in git = branch in Keboola. Diff it, review it, push it back.
Context stays free for your work
Over a hundred operations that the agent looks up as it goes, instead of filling its context up front.
Ten operations, one line
Loop twenty projects, filter, save. One step instead of twenty, and less room to guess.
Get started
Connect your projects.
Any AI agent gets there through the kbagent plugin. Sign in once and every project you can reach comes with it.
Claude Code setup- 01
Open Claude Code
TerminalRun steps in your Terminal session with
clauderunning. Don't have it yet? Follow the Claude Code quickstart. - 02
Add the Keboola marketplace
Points Claude Code at the Keboola plugin catalog. The plugin installs next.
- 03
Install the plugin
Adds Keboola skills, expert agents and slash commands to Claude Code.
We recommend the default user scope, so
kbagentis available every time you open Claude Code, not just in this one folder. - 04
Run the setup
Insert the address you see in the Keboola UI, for example
https://connection.keboola.com. It starts the browser sign-in and registers every project you can reach. You finish the sign-in in your terminal.
Using different AI? All supported.
Log in to your Keboola account, go to Project Settings → Agentic CLI. The setup for your client is there, with the command already filled in for your project.
Everything you need
One CLI. 13 ways to get things done.
From config search to AI permission firewalls, every command outputs structured JSON so agents can parse it, and rich tables so humans can read it.
Capabilities
Configuration management
List, search, inspect, update, delete configs across all connected projects. Full-text search finds that one transformation across 30 projects in seconds.
user@keboola:~/projects $ kbagent config search --query "customer_id"user@keboola:~/projects $ kbagent config new --component-id keboola.snowflake-transformation
How it fits
Four ways in.
One platform.
The CLI is one way to work with Keboola, not a replacement for the others. Use the UI to look around, the CLI or MCP to let agents do the work, the API when you want the primitives.
For agents and humans
Keboola CLI
Every operation as a command an agent can compose, dry-run and drop into CI.
For agents
MCP server
Connect your data to third-party AI tools like Claude.ai, ChatGPT or Make. They query tables, run jobs and build flows, no terminal needed.
How MCP worksFor humans
Web UI
Browse projects, run jobs, debug visually. Ask Kai about your data without leaving the browser.
Meet KaiFor agents and humans
Raw API
Storage, Queue, Encryption, Sandboxes. Every primitive available directly.
API reference
Further reading
