How this works

This app is powered by Jetty, managed infrastructure for agentic AI workflows. When you upload a PDF, Jetty provisions an isolated sandbox, runs an AI agent with a detailed set of instructions called a runbook, and returns the results — all through a single API call.

What happens when you upload a paper

1
Your PDF is uploaded to cloud storage
The file goes directly to Jetty's sandbox storage via a presigned URL — it never touches an intermediate server.
2
Jetty provisions a sandbox
An isolated container is spun up with Python, the mlcroissant library, and a coding agent (currently Claude Code). The agent has full autonomy inside the sandbox: shell, network, file system.
3
The agent follows the runbook
The runbook tells the agent exactly what to do: read the paper, extract dataset metadata (name, description, splits, features, license, citation), build a Croissant JSON-LD file, validate it against the MLCommons schema, and iterate up to 3 times if validation fails.
4
Results come back
The agent writes croissant.json, summary.md, and validation_report.json to the sandbox. Jetty persists every artifact to cloud storage and records a full execution trajectory so you can inspect each step.
5
The sandbox is destroyed
Once the run completes, the container is torn down. Only the output files and trajectory remain.

What is a runbook?

A runbook is a structured Markdown document that tells a coding agent how to accomplish a complex, multi-step task end-to-end. Think of it as a recipe: it defines the objective, the exact files the agent must produce, the steps to follow, and the evaluation criteria that determine whether the output is good enough.

Section	Purpose
Frontmatter	YAML with version, evaluation strategy, agent, model, and sandbox snapshot
Objective	What the agent is doing, what it's producing, and for whom
Required output files	Every file the agent must write — the task isn't complete until they all exist
Parameters	Template variables injected at runtime (e.g. `{{pdf_filename}}`)
Steps	Sequential instructions — each step can run code, call tools, or write files
Evaluation	Programmatic validation or rubric-based scoring to decide if the output passes
Iteration	If outputs fail evaluation, the agent reads the error, applies a fix, and retries (bounded, typically 3 rounds)
Final checklist	Imperative verification to ensure all outputs exist before the run ends

You can view this app's runbook to see exactly what the agent does when it processes your paper.

Run it yourself

You don't need this web app. The runbook is portable — you can run it locally with your own agent or through Jetty's API with any supported agent.

Option A: Run locally with your own agent

1.Copy the runbook from the Runbook page and save it as RUNBOOK.md
2.Place your PDF in the same directory
3.Run with your preferred agent:

Claude Code

claude --system-prompt RUNBOOK.md \
  "Generate a Croissant file for the dataset in paper.pdf"

Codex

codex --system-prompt RUNBOOK.md \
  "Generate a Croissant file for the dataset in paper.pdf"

Gemini CLI

gemini --system-prompt RUNBOOK.md \
  "Generate a Croissant file for the dataset in paper.pdf"

The agent will install dependencies, read the PDF, build the Croissant file, validate it, and iterate until it passes — all on your machine.

Option B: Install the Jetty skill

The Jetty skill connects your coding agent to Jetty's infrastructure. Workflows run in isolated sandboxes with full persistence — no local setup required.

Claude Code

claude mcp add jetty -- npx -y jetty-mcp-server

Cursor / VS Code / Windsurf / Zed

// Add to your editor's MCP config:
{
  "mcpServers": {
    "jetty": {
      "command": "npx",
      "args": ["-y", "jetty-mcp-server"],
      "env": {
        "JETTY_API_TOKEN": "mlc_your_token_here"
      }
    }
  }
}

Once installed, run /jetty-setup in Claude Code to walk through account creation and API key configuration.

Option C: Call the Jetty API directly

Jetty exposes an OpenAI-compatible /v1/chat/completions endpoint. Send a runbook as the system prompt with a jetty block, and Jetty handles sandbox provisioning, agent execution, and artifact persistence.

curl -X POST https://flows-api.jetty.io/v1/chat/completions \
  -H "Authorization: Bearer $JETTY_API_TOKEN" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "claude-sonnet-4-6",
    "messages": [
      {"role": "system", "content": "<your runbook>"},
      {"role": "user", "content": "Generate a Croissant file for paper.pdf"}
    ],
    "jetty": {
      "runbook": true,
      "collection": "my-collection",
      "task": "pdf2mlcroissant",
      "snapshot": "python312-uv",
      "file_paths": ["path/to/paper.pdf"]
    }
  }'

Supported agents

Agent	Default model	API key needed
Claude Code	claude-sonnet-4-6	ANTHROPIC_API_KEY
Codex	gpt-5.4	OPENAI_API_KEY
Gemini CLI	gemini-3.1-pro-preview	GOOGLE_API_KEY

Any MCP-compatible agent (Cursor, VS Code Copilot, Windsurf, Zed) can also connect to Jetty via the MCP server.

Learn more

Jetty documentation — full platform docs, guides, and API reference
Writing runbooks — how to write your own runbooks from scratch
Try Jetty — sign up and get your API token