build · orchestrate

Wire up a team of agents, visually.

Drop a Planner, Coder, Reviewer, Tester, or Adversary onto a canvas, wire them in order, and run the whole pipeline. They hand off one to the next — or fan out as a swarm — and every node lands as a span in the trace.

Start free Open Canvas →

canvas · pr-reviewrunning

Plannerdone

Coderlive

Reviewerqueued

agent.run · 3 spans2 / 3 nodes · streaming

Five role-based nodesSequential hand-off or swarmLive token-by-token transcript

Composition without code

Seed a pipeline from a starter template — Planner → Coder → Reviewer is the canonical one — then add, remove, and reorder nodes. Each role ships with a sharp default system prompt, and you can bind any node to a saved agent from your roster to pin its model and instructions. No graph DSL, no YAML.

canvas · pr-reviewrunning

Plannerdone

Coderlive

Reviewerqueued

agent.run · 3 spans2 / 3 nodes · streaming

Hand-off or swarm

Run sequentially and each agent sees the full chain so far: the output of every prior step is folded into the next node's system prompt as a labeled block, so the Reviewer reads what the Coder actually wrote. Or dispatch the same task to every node at once as a swarm and compare their answers side by side.

missions · board4 lanes

Backlog2

Migrate auth to magic link

P1Ship Squad

Draft changelog entry

In progress1

Refactor billing webhook

Ship Squadrunning

Done1

Add OpenAPI spec

Docs CrewTrace

Archived1

Spike: vector search

@all ship the billing fixSend

Watch it think, live

Runs stream token-by-token into a transcript over a private Realtime channel, batched for smoothness. You see each node start, emit, and complete in real time — no refreshing, no polling, no waiting for the whole pipeline to finish before you know it went sideways.

trace · run_8c4fok · 742ms · $0.0053

agent.plan742ms

tools.search_code86ms

llm.complete_refactor612ms

tools.edit_file78ms

evals.no_regression54ms

agentllmtooleval5 spans · 3,007 tok

The pipeline is the trace

Every canvas run is one agent.run root span with an llm.completion child span per node — carrying its model, token counts, cost, and stop reason. The visual pipeline and its observability are the same object: open the trace from the run and replay exactly what each agent saw and said.

trace · run_8c4fok · 742ms · $0.0053

agent.plan742ms

tools.search_code86ms

llm.complete_refactor612ms

tools.edit_file78ms

evals.no_regression54ms

agentllmtooleval5 spans · 3,007 tok

5 + custom

Role nodes

chain · swarm

Run modes

1 OTel span