Skip to main content

Claude Code Plugin

The Jetty plugin for Claude Code gives you two slash commands — /jetty-setup for first-time onboarding and /jetty for managing workflows with natural language.

Install

claude plugin marketplace add jettyio/agent-skill
claude plugin install jetty@jetty

First-Time Setup (/jetty-setup)

Run /jetty-setup and the plugin will walk you through:

  1. Account — Sign up or log in at flows.jetty.io
  2. API token — Create a token at Settings → API Tokens and paste it in
  3. Provider API key — Choose OpenAI or Google Gemini and store the key in your collection
  4. Demo workflow — Deploys and runs the "Cute Feline Detector" (image generation + LLM-as-judge)

The setup takes about 3–5 minutes and leaves you with a working collection and a deployed workflow.

Where the Token Goes

The plugin stores your token in your project's CLAUDE.md:

I have a production jetty api token mlc_xxxxxxxxxxxxx

This lets the agent use the token automatically in future sessions.

warning

If your CLAUDE.md is checked into git, add it to .gitignore or use a project-level .claude/settings.json instead. API tokens should not be committed to version control.

Managing Workflows (/jetty)

After setup, use /jetty followed by what you want to do:

/jetty list my tasks
/jetty create a workflow that compares GPT-4o and Claude on summarization
/jetty run cute-feline-detector with prompt "a penguin wearing a top hat"
/jetty show me the last 5 runs of cute-feline-detector
/jetty what step templates are available for image generation?

The /jetty command has access to all 14 MCP tools under the hood — it can create tasks, run workflows, inspect trajectories, add labels, and more.

Tips

  • Token inline: The agent always inlines the token in curl commands (TOKEN="mlc_..." then $TOKEN). Environment variables don't persist across shell invocations in Claude Code.
  • Two APIs: The agent routes to the right API automatically — dock.jetty.io for collections/tasks, flows-api.jetty.io for running workflows and viewing results.
  • Batch scripts: For complex multi-step bash commands, the agent writes to /tmp/script.sh and runs it — this avoids zsh escaping issues.
  • Polling: Workflows take 30–45 seconds. The agent waits, then polls trajectories for results.

What's Happening Under the Hood

The plugin uses Jetty's MCP server, which provides these tools:

ToolPurpose
list-collectionsList your collections
list-tasks / get-taskBrowse and inspect tasks
create-task / update-taskCreate or modify workflow definitions
run-workflow / run-workflow-syncExecute workflows (async or sync)
list-trajectories / get-trajectoryView workflow run results
get-statsExecution statistics
add-labelTag trajectories with labels
list-step-templates / get-step-templateBrowse available activities

For details on each tool, see the MCP Server page.

Next Steps