Open source · MIT

The canvas for agents.

Place nodes on an infinite canvas, wire them top‑to‑bottom, and press Run. Each node hands its output — and control — downstream: investigate, plan, design, review — executing as one chain you watch in real time.

View on GitHub See the canvas
MIT licensed· OpenAI · Anthropic · Google· Self-host or local
any model, per node
OpenAI Anthropic Google + any compatible endpoint
why a canvas

Stop prompting in a loop.

A chat window makes you the runtime — prompt, wait, copy, paste, re-prompt. DISPATCH lets you wire the work once and press Run. Every messy turn on the left becomes one deterministic node on the right.

prompting the old way

You babysit every step.

you

Research our top 3 competitors and gather the latest numbers.

ai

Sure — which competitors did you have in mind?

you

The three from the deck. Now compare their pricing tiers.

ai

Here's a comparison. (copy → paste → re-read…)

you

Good. Draft a one-page summary from that.

you

Now check it for errors before I send it.

the canvas with dispatch

You wire it once. It runs itself.

Run ready
Investigateclaude · sonnet
Research the 3 competitors from Context → deck.
web_search → 3 sources · no "which ones?" — pre-scoped
Compareopenai · gpt-4o
Build a pricing-tier table from the findings.
edge carries output in — no copy / paste
Draftclaude · opus
Write a one-page summary from the table.
summary.md · 412 words
Reviewerror pass · gate
Proof for errors, then hold for sign-off.
2 fixes · awaiting approve

Four turns of babysitting → four wired nodes. Press Run once.

the canvas

One graph. Every capability, live.

Each tile is a real DISPATCH.AI surface — not a screenshot. Move across the grid; the canvas reacts.

live run idle
Initialisertrigger
Investigateweb_search
Reviewhuman checkpoint
model hotswap

Swap provider live.

Create gpt-4o
live tools
› web_search
› fetch_url
› read_file
› write_file
›
context + memory

Shared across the graph.

Context run.findings write·read
Context run.plan write·read
parallel + merge Parallel Design Doc
Merge fan-in → 1 join
real-time status
idle running done paused error

Every node reports state as it runs. Edges animate while data flows downstream.

analytics · node-delegated cost this run

Each node meters its own tokens and spend. The canvas rolls every node up into the dashboard total — per-node accounting on the left, per-model on the right.

per nodecost
Investigate12.4k tok$0.04
Plan6.1k tok$0.02
Create38.7k tok$0.11
Design9.8k tok$0.05
Review2.3k tok$0.01
rolled up69.3k tok$0.23
per modelshare
gpt-4o$0.13
claude-opus$0.07
llama-70b · local$0.03
Open analytics
agentic streaming

Deploy 20 agents on one task.

Fan a single task out across twenty agents. They run in parallel, stream their work live, and converge into one synthesized result — minutes, not hours.

20 agents · 1 task · parallel · minutes, not hours
parallel fan-out idle
Refactor & harden the auth layer task
Synthesized result merge · 20 → 1
How fan-out, streaming & synthesis work open feature
everything, in one place

Nineteen capabilities. One canvas.

Every feature is a node, a surface, or a switch on the same graph. Open any tile to go deeper.

orchestration
Activeflagship

Agentic Streaming

Deploy 20 agents on one task — parallel, streamed, synthesized.

open
In developmentcontrol

Advanced Orchestration

Branch, retry, cap concurrency, loop, nest, gate, timeout.

open
Testingplanning

Canvas Planning

Lay out goals, phases, and milestones right on the graph.

open
Planningassertions

Testing Pipeline Nodes

Drop test nodes into the chain and assert on any output.

open
models
Activeper node

Model Hotswapping

Set provider + model per node and swap live, mid-run.

open
In developmentoffline

Local Models

Ollama, llama.cpp, LM Studio — fully local, alongside the cloud.

open
In developmentMODELS.md

Directive Model Control

One version-controlled file governs every model on the canvas.

open
Testingvia MCP

Subscription Models

Wire non-API subscription models straight into nodes via MCP.

open
knowledge + extensibility
Planningretrieval

Node-based RAG

Embed → search → top-k → context, wired in like any other node.

open
Activecommunity

Skills-as-Nodes

Publish a skill; it becomes a node anyone can drop on the canvas.

open
PlanningMCP

MCP + Integrations

Roblox, Linear, Notion, Vercel, Safentic, Netlify — and more.

open
Testingprompt → graph

Model Prompt

Model-driven prompting that generates whole node graphs — its own live read of the canvas.

open
interface + control
ActiveVoiceOrb

Agentic Voice Control

Talk to your canvas — it edits the graph as you describe it.

open
ActiveCLI

Inbuilt Terminal

run, stop, retry, stream logs — a real CLI inside the canvas.

open
ActiveLAN

Local-network Collaboration

Name-tagged cursors, inline comments, live node-wiring — no cloud.

open
Planningusage

Analytics

Tokens and spend across every provider and local model.

open
infrastructure
PlanningRunPod GPU

Rapid Pod Deployments

Ship a canvas to a GPU pod serving a 70B-class model.

open
Activelocal-first

Secure-first / Fully Local

Self-hosted, no telemetry, no cloud — air-gap friendly.

open
In developmentversioned

Git-backed Canvas

Every workflow is a file — commit, diff, and review like code.

open
MIT

Open source · build your own

Fork the runtime, write a skill, add a node type. Yours to own.

view source
local-network collaboration

Your whole team, one live canvas.

Real-time over your LAN. Name-tagged cursors, inline comments, and live node-wiring stream instantly between everyone. No cloud, no account, nothing leaves your network.

named cursors ·inline comments ·live wiring ·local-first
Explore live collaboration
shared canvas · lan 4 connected
Investigate web_search
Plan opus
Create draft
Review gate
ana
kai
sam
you
peers: ana · kai · sam · you · 192.168.1.0/24 · syncing
model prompt Testing

Describe it. Watch the graph build itself.

Model-driven prompting that generates a whole node graph from a single prompt — prompt → graph, fully model-controlled. It keeps its own live read of the canvas: deciding which nodes to place, how to wire them, and which model and tools each node gets — adapting as the graph grows.

prompt → graph ·model-controlled ·live canvas read ·self-assembling
prompt → graph ready
your prompt
›
model reasoning
See how Model Prompt works
canvas · model-built 0 nodes · 0 edges
voice control

Talk to your canvas.

A particle VoiceOrb listens, thinks, and speaks — editing the graph as you describe it. Watch it react.

VoiceOrb state: idle — a gentle breathing sphere.

The VoiceOrb has four states. Idle: a gentle breathing sphere. Listening: the surface ripples toward your cursor as it captures speech. Thinking: turbulent high-frequency motion while it plans the edit. Speaking: rhythmic pulses in blue and accent as it responds and applies the change.

transcript
user

dispatch

Added a Review node after Investigate and wired the edge.

Investigate
Review + added
Explore voice control
rapid pod deployments upcoming flagship

Ship your orchestration to a GPU. Out of the box.

Deploy a whole canvas onto a GPU pod with one command. Via RunPod, DISPATCH provisions the hardware and serves a multi-billion-parameter, 70B-class model out of the box — your chain runs in minutes, no infra to wire up. Landing soon.

runpod ·GPU pod ·70B served ·out of the box
deploy provisioning
$ dispatch deploy --pod
canvas packageddone
runpod · A100 80GB acquireddone
serving 70B modellive
endpoint → orchestrationqueued
See how pod deployment works
pod
GPU pod runpod
gpuA100 80GB
model70B-class
providerrunpod
statusserved

pod: A100 80GB · runpod · 70B served

Built by one 17‑year‑old.
Open‑sourced so it can go further.

One week into an accelerator, ~$250 in API credits deep into testing, it clicked: DISPATCH can redefine the messy part of agentic orchestration — where many agents hand off, gate, and converge on real work. Testing that at API prices is expensive, so it's MIT and open. Any help genuinely moves it forward.

View on GitHub Get started