Build agents on a canvas.
Run, rewind, fork.
Compose multi-agent workflows visually, watch tokens stream through every node, then scrub back to any checkpoint, edit the state and fork a new run from that step. Evals, model sweeps and MCP tools are built in.
Debug agents like code, not like black boxes.
Console, timeline, cost and traces all derive from one event log, so what you see on the canvas is what actually ran.
Rewind any run, fork from any step
Every node completion is a checkpoint. Scrub back, see what changed, edit a variable and launch a new run from that point. The original stays untouched; compare the two side by side.
- after Write draftpaused
- edited by handfork
- after Rewritedone
Evals that point at the node
Build a dataset, pick scorers, run it. Failures light up the node that wrote the bad value. Sweep candidate models for one node and apply the cheapest that still passes.
Humans and tools in the loop
Human nodes pause the run with a card that shows exactly what to approve. Tool nodes call built-ins or any MCP server you connect. Code nodes run Python in a Docker sandbox.
Drag nodes, connect ports, edit prompts in Monaco with {{variables}}.
Tokens render inside the node that produces them, over a WebSocket.
A waterfall of node spans and a filterable event log for every run.
Metrics, steps and final variables diffed side by side.
Connect a server over stdio or HTTP; its tools appear in the Tool node.
Per-node calls, tokens, p50 and p95 latency and spend, priced per model.
What ran here this week
Runs, last 7 days
0 runs · $0.00000Recently edited
All graphs- No graphs yet.
Run outcomes
0 runs- no runs yet
No model keys configured. Add a Groq or Gemini key to the API.
Start from a template
- 1ComposeDrag nodes onto the canvas and wire them up.
- 2RunPress ⌘↵ and watch tokens stream through each node.
- 3RewindScrub checkpoints, edit state, fork from any step.
- 4EvalScore a dataset, find the cheapest model that passes.
See a run rewind in under a minute.
Open the tweet-writer template, run it, answer the human gate, then fork from the draft step. The whole loop costs a fraction of a cent on Groq's free tier.
6 nodes · model, human gate, router · about 2 model calls per run