Skip to main content

Jetty

One API for AI workflows, evaluation, and agentic execution.

Jetty gives your team a single /v1/chat/completions endpoint that does three things: proxies 100+ LLM providers, orchestrates multi-step evaluation pipelines, and runs autonomous agents in isolated sandboxes. Drop it into any OpenAI-compatible integration — existing SDKs work out of the box.

Start here​

Chat Completions API​

One endpoint, two modes. Without the jetty block it's a standard LLM proxy with automatic trajectory recording. With the jetty block it provisions a sandbox, runs an agent, and returns structured results.

Works with any OpenAI SDK. Switch providers by changing the model field.

Read the API reference →

Architecture​

Understand how Jetty's three engines — Passthrough, Workflow, and Runbook — connect through a single API layer to persistence, tracing, and object storage. See how Collections, Tasks, and Trajectories organize your work.

See the architecture →

Build with Jetty​

Writing Runbooks​

A runbook is a structured markdown file that tells a coding agent how to accomplish a complex task end-to-end — with evaluation loops, iteration, and quality gates. When the first attempt is rarely sufficient, a runbook encodes your domain expertise into a repeatable process.

Covers the canonical structure, frontmatter schema, evaluation patterns, common pitfalls, and the /create-runbook wizard.

Learn to write runbooks →

Agent Skill & MCP​

Use Jetty directly from Claude Code, Cursor, Windsurf, VS Code Copilot, Zed, or Gemini CLI. Your agent can create workflows, kick off runs, inspect trajectories, run evaluations, and browse 40+ step templates — all without leaving your editor.

Three connection methods: Claude Code plugin (/jetty commands), MCP server, or raw REST API.

Set up agent integration →

Guides​

Hands-on tutorials for real workflows:

Browse all guides →

Quick Start​

Get running fast with progressive tutorials:

  • 60 seconds — Generate an image in one API call.
  • Setup — Get your API token and configure model keys.
  • First flow — Create and run your first workflow.
  • Model comparison — Compare GPT-4, Claude, and Gemini side-by-side (5 min).
  • Agent benchmarking — Test coding agents with TerminalBench (10 min).

Start the quick start →

How it works​

# Passthrough mode — standard LLM proxy, every call recorded as a trajectory
curl https://flows-api.jetty.io/v1/chat/completions \
  -H "Authorization: Bearer $JETTY_API_TOKEN" \
  -d '{
    "model": "claude-sonnet-4-6",
    "messages": [{"role": "user", "content": "Summarize this quarter'\''s metrics"}],
    "stream": true
  }'

# Runbook mode — add a jetty block to run an agent in an isolated sandbox
curl https://flows-api.jetty.io/v1/chat/completions \
  -H "Authorization: Bearer $JETTY_API_TOKEN" \
  -d '{
    "model": "claude-sonnet-4-6",
    "messages": [{"role": "user", "content": "Run the evaluation suite"}],
    "jetty": {
      "collection": "my-evals",
      "runbook_url": "https://raw.githubusercontent.com/org/repo/main/RUNBOOK.md"
    }
  }'

Both modes return trajectories — full execution traces with inputs, outputs, and metadata — so you always have observability into what happened and why.

More resources​

Step Library 47+ pre-built activities: AI models, control flow, data processing, evaluation.

Browse steps →

Examples Copy-paste workflow JSON for common tasks: chat, image gen, batch processing, translation.

View examples →

API Reference Chat completions, webhooks, GitHub PR integration, and authentication.

Read the API docs →