interface + control

Inbuilt Terminal / CLI

A real terminal, docked in the canvas. Type run, retry Review, stream 20 or deploy — the graph reacts. Drive every node without leaving the keyboard.

Active Open source · MIT keyboard-first

what it is

The canvas, but you type to it.

The graph and the terminal are two views of one runtime. Anything you can do by dragging and wiring, you can do with a verb and a node name — and the prompt talks back: states, logs, costs, deploy URLs.

No shell to context-switch to, no IDs to copy. The CLI is the orchestration layer — the same engine behind the Run button, exposed as a language.

the verbs dispatch >

run

Execute the graph top-to-bottom from the Initialiser.

stop

Halt the run; in-flight nodes pause, edges freeze.

retry <node>

Re-run one node — e.g. retry Review — keeping upstream context.

model <node>

Hotswap a node's provider + model in place.

stream <n>

Fan a task out across n parallel agents.

deploy

Package the canvas and ship it to a pod; prints the URL.

install <skill>

Pull a community skill; it becomes a droppable node.

voice

Hand control to the VoiceOrb and speak the graph.

status · logs

Per-node state, tokens, spend; tail a node's live log.

save · export

Commit the canvas to a file; export as a portable prompt.

try it

Type a command. Watch the graph move.

Pick a command below. The terminal echoes it, the runtime responds, and the canvas on the right updates node states, edges and the deploy readout in real time.

dispatch — terminal idle

DISPATCH.AI terminal · type a command above, or press run

graph: 4 nodes · Initialiser → Investigate → Create → Review

dispatch >

canvas 4 nodes

Initialisertrigger

Investigateweb_search

Creategpt-4o

Reviewgate

pod: not deployed

Every command maps 1:1 to the real terminal — same grammar, same node engine.

how it works

One runtime, two surfaces.

The terminal isn't a wrapper that scripts the UI — it speaks straight to the orchestrator. So does the canvas. They stay in lock-step.

parse

Verb + node, nothing more.

Commands resolve to node IDs on the active graph. retry Review targets the Review node; tab-completion knows every node and skill by name.

dispatch

Same engine as Run.

The parsed command is handed to the orchestrator that drives Parallel, Merge, Apply and Deploy. No second code path — the terminal can't drift from the canvas.

stream back

The prompt talks back.

State transitions, per-node tokens, logs tails and deploy URLs print inline — the same events that pulse the dots and march the edges.

scriptable + reproducible

A run you can paste.

Because every action is a line, a whole workflow becomes a short, reviewable script. Drop it in a README, share it in an issue, replay it on any canvas.

$ model Create anthropic claude-opus

$ run

$ retry Review

$ deploy → dispatch.run/auth-harden

drives every node type

Nothing is off-limits.

If it's a node, the CLI can address it — including control-flow and infra nodes.

InitialiserInvestigatePlan DesignCreateEvaluate DocApplyContext ReviewParallel MergeDeploy

Back to all features

related features

ActiveVoiceOrb