Z.E.N.

Appearance

Dispatch from your agent ​

You probably already live in an AI tool. Claude Code, OpenClaw, Hermes, ChatGPT, something with a chat surface and a memory of what you've asked it. Z.E.N. is designed to sit behind that tool, not replace it. Your agent calls Z.E.N. workflows by name; the response comes back into your existing chat.

The shape ​

Your agent learns about a workflow once. Either you tell it the workflow name, or it discovers the workflow on its own by reading ~/.zen/workflows/ or hitting the local Z.E.N. API. From then on, you ask the agent the way you'd ask any other tool: "run morning-brief," "fire the research sprint for Q3 planning," "kick off the content draft pipeline." The agent calls Z.E.N., the workflow runs, the output streams back into your chat.

This is the loop that closes the homepage feature called "Inside your agent." Z.E.N. is the structured-work infrastructure your agent reaches for; the agent stays the place you actually do work.

Setup: Claude Code ​

Claude Code reads ~/.zen/workflows/ automatically when Z.E.N. is installed. Workflows show up as tools the model can call. To fire one, ask Claude:

Run morning-brief.

Claude calls Z.E.N. via the local CLI, waits for the run to finish, and quotes the output back. If the workflow has a schedule, Claude can ask Z.E.N. to enable, disable, or change it.

To make Claude Code aware of a new workflow you just wrote, restart the Claude Code session or ask it to "reload zen workflows."

Setup: OpenClaw ​

OpenClaw treats Z.E.N. as a tool surface via the local HTTP API. Add the Z.E.N. integration in OpenClaw's settings; point it at http://localhost:3090 (the default daemon port). OpenClaw will see every registered workflow and let you fire it by name.

If you don't run the daemon, OpenClaw can fall back to the CLI dispatch path. The setup screen has both options.

Setup: Hermes ​

Hermes uses the same MCP-style dispatch as OpenClaw. Configure the Z.E.N. MCP server (it ships with the install) in Hermes' MCP config, restart Hermes, and your workflows are callable from inside any Hermes conversation.

Setup: any other agent ​

The general pattern is the same. If your agent supports tool calls or MCP servers, point it at Z.E.N.'s local HTTP API or the CLI. If it doesn't, you can still fire workflows by asking your agent to run a shell command (zen workflow run <name>).

The full HTTP API is at Reference / API. The MCP server install path is at MCP servers.

What your agent can do beyond firing ​

A connected agent can:

  • List your workflows and what they do.
  • Start a fire and watch the run stream back.
  • Cancel a run that's still going.
  • Enable, disable, or modify a schedule.
  • Run the workflow builder ("describe a workflow, get a YAML file") from inside your chat.

The agent doesn't replace the Web UI or the CLI; it gives you a way to do all of that without switching apps.

Picking a provider ​

Separate question, often confused with this one. The provider is the AI that runs a single node inside a workflow. Z.E.N. ships with Claude, Codex, and Pi as built-in providers. The agent that calls Z.E.N. is a different thing from the provider that runs the steps; you can dispatch from Claude Code and have a step inside the workflow use Codex if that's what the work needs.

Set the per-node provider in the workflow YAML:

yaml
nodes:
  - id: summarize
    type: prompt
    provider: codex
    prompt: |
      ...

Or set a default for the workflow under assistants:. See Configuration for the full shape.

Next ​

AI that follows a recipe, not a conversation.

Z.E.N.