Choose How You Use AI with Kestra

Compare every way to use AI with Kestra and pick the one that fits how you work.

AI shows up in Kestra in two unrelated ways, and it helps to keep them apart:

• AI that helps you build and operate Kestra — writing flow YAML, fixing a failed execution, deploying to a namespace. The AI is a way to drive Kestra faster; it never runs as part of your workflow.
• AI that runs inside your flows — a task calls a model, an agent decides what to do next, a flow answers questions over your documents. Here the AI is part of what you orchestrate.

None of this is required. Kestra is an orchestrator, and every flow you can generate with AI you can also write by hand in YAML or build in the No-Code editor.

Build and operate Kestra with AI

Start from where you already work:

AI Copilot in the Kestra UI

AI Copilot is built into the Kestra editor. Describe what you want — for example, “make a REST API call to https://kestra.io/api/mock and allow failure” — and Copilot generates the flow YAML for you to review, accept, or edit. It also modifies existing flows incrementally, touching only the part you asked about, and powers Fix with AI on failed tasks.

Copilot is the fastest path when you are already in the UI and want the change applied to the flow in front of you.

Your own coding agent

If your flows live in Git next to your application code, you may prefer to stay in Claude Code, Cursor, Codex, or Windsurf. Three separate pieces make that work, and each covers a different gap:

• MCP server — context. A hosted Model Context Protocol endpoint that gives your agent live access to task schemas for all 1,400+ plugins, the blueprints library, and documentation search. Without it, an agent invents plugin properties from stale training data.
• Agent Skills — know-how. SKILL.md files that teach your agent how to author valid flows (kestra-flow), harden them for production (kestra-flow-hardening), operate an environment (kestra-ops), and port Airflow DAGs (migrate-airflow-kestra). Skills carry the guardrails: no invented task types, no hardcoded secrets, validate before deploy.
• kestractl — hands. The Kestra CLI the agent actually calls to validate, deploy, and run flows against a real instance. The kestra-ops skill is a wrapper around these commands, so an agent without kestractl can write flows but cannot deploy them.

Set it up

Connect the MCP server. For Claude Code:

claude mcp add kestra --transport http https://api.kestra.io/v1/mcp

Any MCP-compatible tool can connect to https://api.kestra.io/v1/mcp over HTTP — see MCP server for Cursor and other configurations.

Install the skills. skills.sh detects your agent and writes the files to the right location:

npx skills add kestra-io/agent-skills

Install kestractl and point it at your instance:

curl -fsSL https://raw.githubusercontent.com/kestra-io/kestractl/main/install-scripts/install.sh | bash
kestractl config add default http://localhost:8080 main --username YOUR_USERNAME --password YOUR_PASSWORD --default

See kestractl for Enterprise token auth and the full command reference.

What it looks like end to end

With all three in place, a single prompt covers authoring and deployment:

Use kestra-flow to write a flow in namespace company.data that fetches
https://api.example.com/metrics every 30 minutes and stores the response in
the KV store under the key last_metrics. Validate it with kestractl, then
deploy it to the dev instance.

The agent queries the MCP server for the correct plugin schema, follows the skill’s guardrails (no hardcoded credentials, valid YAML structure), runs kestractl flow validate, and deploys with kestractl flow create.

Expose flows as AI tools

MCP Server lets you go the other direction: instead of an agent driving Kestra, Kestra flows become callable tools for any MCP-compatible AI agent. Add an McpToolTrigger to any flow and it is automatically registered as a named tool. AI agents such as Claude Desktop, Claude Code, and Cursor can then discover and invoke your flows directly, with inputs and outputs mapped to a JSON schema.

Run AI inside your flows

• AI Agents: autonomous orchestration where an LLM decides which actions to take at runtime rather than following a fixed sequence.
• AI Workflows: pre-built flow templates for common AI patterns such as summarization, classification, and structured extraction.
• RAG Workflows: retrieval-augmented generation pipelines that connect your documents to a language model.