Kestra MCP: A Live Plugin and Blueprint Catalog for AI Coding Agents

Writing a Kestra flow from scratch means knowing what plugins exist, which tasks they expose, and what properties each task accepts. For simple, familiar workflows this is fine. For anything more complex — choosing the right S3 trigger variant, checking what secret managers are supported, or confirming whether a Blueprint already covers your use case — you end up context-switching to the docs browser mid-task.

The Kestra MCP server eliminates that friction. It is a remote HTTP server at https://api.kestra.io/v1/mcp that speaks the Model Context Protocol, exposing the full Kestra plugin registry and Blueprint catalog as callable tools. Connect it to your AI coding agent once; from that point, the agent can discover plugins, inspect task schemas, and fetch Blueprint YAML without ever leaving the conversation.

What is MCP?

The Model Context Protocol is an open standard for connecting AI agents to external tools over a well-defined interface. An MCP server declares a set of tools; a compatible client (your AI agent) invokes them by name and receives structured results. Because MCP is transport-agnostic, the Kestra server uses HTTP with Server-Sent Events — no local process to install or manage, just a URL.

Connecting Claude Code

Add the Kestra MCP server to Claude Code with a single command:

# Project-scoped (stores in .claude/settings.json in current repo)
claude mcp add --transport http kestra https://api.kestra.io/v1/mcp

# User-scoped (global, available in every project)
claude mcp add --transport http --scope user kestra https://api.kestra.io/v1/mcp

Or add it manually to .claude/settings.json:

{
  "mcpServers": {
    "kestra": {
      "type": "http",
      "url": "https://api.kestra.io/v1/mcp"
    }
  }
}

Once connected, the 9 tools are immediately available in every Claude Code session. You can verify with /mcp, or simply start a flow-writing task — Claude will invoke the relevant tools automatically.

What the server exposes

All 9 tools carry readOnlyHint: true. The server is a pure catalog; it never modifies state.

Plugin discovery

Task documentation

Blueprints

Plugin release history

Full-text search

Use cases

Writing a flow from scratch

Ask Claude Code to write a flow that reads from S3 and loads into BigQuery. Instead of guessing task names or hallucinating property shapes, the agent calls list_plugins to discover plugin-aws and plugin-gcp, then plugin_tasks for each to enumerate available tasks, then task_schema on io.kestra.plugin.aws.s3.Downloads and io.kestra.plugin.gcp.bigquery.Load to get exact property names, types, and defaults. The resulting YAML is valid on the first attempt.

Checking what plugins are available

“Does Kestra support Vault for secrets?” — the agent calls list_plugin_elements(elementType: "SECRET_MANAGER") and returns the full list including the FQCN and description for the HashiCorp Vault implementation. The same pattern works for TASK_RUNNER (“can I run tasks on GCP Batch?”), STORAGE (“does Kestra support MinIO?”), and TRIGGER (“is there a Kafka realtime trigger?”). No guessing, no docs tab, no stale answers.

Bootstrapping with Blueprints

Before writing a flow from scratch, the agent calls search_blueprints with a query like "dbt" or "slack notification", scans the matches, and calls get_blueprint_flow on the best result to retrieve the full YAML. This is faster than authoring from scratch and guarantees the output follows Kestra’s own patterns and best practices.

Investigating plugin compatibility

When a flow breaks after a plugin upgrade, the agent calls plugin_release_history to retrieve the full release history with the target Kestra version for each tag. This surfaces the exact release where a change was introduced without leaving the conversation or opening GitHub.

Connecting other AI clients

The same endpoint works with any MCP-compatible tool:

Codex CLI — ~/.codex/config.toml:

[[mcp_servers]]
name = "kestra"
url  = "https://api.kestra.io/v1/mcp"

Gemini CLI — ~/.gemini/settings.json:

{
  "mcpServers": {
    "kestra": {
      "httpUrl": "https://api.kestra.io/v1/mcp"
    }
  }
}

OpenCode — ~/.config/opencode/config.json:

{
  "mcp": {
    "kestra": {
      "type": "remote",
      "url": "https://api.kestra.io/v1/mcp"
    }
  }
}

Why a remote server?

Most MCP servers are local processes: install them, keep them updated, configure them per machine. A remote HTTP MCP server requires none of that. There is nothing to install, nothing to update locally, and no background process consuming memory. The catalog stays current because it mirrors the live plugin registry and Blueprint index that powers kestra.io/plugins — updated on every Kestra release.

Frequently asked questions

What is the Kestra MCP server?

A remote HTTP endpoint at https://api.kestra.io/v1/mcp that implements the Model Context Protocol. It gives AI coding agents live access to Kestra’s plugin registry and Blueprint catalog through 9 read-only tools — with no installation required.

How do I connect Claude Code to the Kestra MCP server?

Run claude mcp add --transport http kestra https://api.kestra.io/v1/mcp for project scope, or add --scope user for global availability. You can also add the mcpServers entry directly to .claude/settings.json.

What tools does the Kestra MCP server expose?

9 tools covering plugin discovery (list_plugins, plugin_tasks, list_installed_versions), element-type listing (list_plugin_elements, parameterized by elementType), task documentation (task_schema), Blueprint search and retrieval (search_blueprints, get_blueprint_flow), release history (plugin_release_history), and full-text search (search). The search_blueprints tool accepts a types parameter (comma-separated FQCNs) to filter templates by included task type — combine it with list_plugin_elements to get valid values.

Does the Kestra MCP server require authentication?

No. The endpoint is public and requires no API key or login. All tools are read-only.

Can I use it with AI agents other than Claude Code?

Yes. Codex CLI, Gemini CLI, OpenCode, or any framework that supports MCP over HTTP. Configuration snippets for each are in the post above.