Start from a template
If your code has breakpoints, why does not your business? Step through a delivery that slipped, a funnel that leaks or an agent that fabricates — the same way you would step through code. Star one to pin it here.
Get started
A graph here is an agent loop you can step through: agent → tool → critic → retry → gate → output. Everything runs against your own models and keys — nothing is proxied through us, and the governor runs on this machine.
Add a model
Open the model registry and paste a key, or point it at Ollama or any OpenAI-compatible endpoint. A model with no key is marked amber on the node rather than failing mid-run, so you can see it before you press Run.
Open a template
Start from one of the 37 rather than an empty canvas — the prompts are the template, and a working loop is easier to change than to invent. Code review cycle is the shortest way to see the shape.
Run it, then break it
Press Run and watch the Runs tab stream node by node. Then set a breakpoint and run again: execution pauses before that node, so you can read what it was about to send.
Sim mode spends nothing
Sim produces a synthetic run with no model calls — good for learning the shape. It also never invokes the guard, so a block you are trying to demonstrate will not happen until you switch to Live.
The guard runs before the call
Prompts and completions are classified before anything is sent, so a blocked node costs zero tokens. A block halts its own branch; the rest of the graph carries on.
Loops stop when they stop working
A critic loop that stops changing the draft is cut early rather than running to its iteration limit. A loop still improving is left alone.
Breakpoints
F9 on a selected node, the same key your editor uses. Then step, continue, or stop from the debug bar.
The graph is a file
In the extension a graph round-trips to a .loop.yaml next to your
code, so it reviews in the same pull request and reverts with the same
git revert.
Every run exports
The Runs tab has an audit report: provenance, cost, every governance decision and the full log, print-ready. Estimated figures are labelled as estimates.
Something stopped the sandbox document from rendering here.
It ships as fleet.html beside this page. Opening it in its
own tab uses a different path and usually works when this one does not.