INPUT
Operator request
A new exception arrives from the work queue.
request100% OPEN-SOURCE VISUAL RUNTIME
ComfyUI
for AI agents.
Design the context window. Draw the loop. Control every tool, permission, and handoff. Inspect every turn, then publish anywhere.
LOCAL-FIRSTYour machine, your models, your data.
MODEL-AGNOSTICBring an API key or an existing subscription.
VIBE-CODEABLEEvery project speaks CLI and MCP.
aux — 001
Prompt the agent into existence. Open the graph. Edit every node. Run locally, publish with one click, or self-host the whole thing.
ONE PLANE. THE ENTIRE AGENT.
Stop building
the invisible.
Every tab is a programmable plane. Start with a prompt, open the generated graph, then change any node, edge, context rule, or execution boundary yourself.
CONTEXT
Mount operating context
Policy, history, files, and the current record.
18.2k / 64kAGENT
Resolve the exception
Investigate, decide, and prepare the next safe action.
claude · 8 turnsTOOLS
Controlled workspace
Browser, shell, files, and domain APIs.
4 allowedHUMAN
Approve external action
Pause before any change leaves the workspace.
required01 / PROMPT IT02 / OPEN THE GRAPH03 / RUN & INSPECT04 / PUBLISH OR SELF-HOST
WHY IT EXISTS
Software was built from functions.
Agent software is built from turns.
Coding agents are excellent at software development. But a startup building an agent for claims, logistics, healthcare, finance, or any other vertical still has to rebuild the actual agent system.
Context assembly. Tool loops. Credentials. Permissions. Human review. Compaction. Logs. Cost controls. Automation. Deployment. aux turns that infrastructure into a visible, portable, and programmable environment.
01CONTEXT
Shape what the model sees.
Build context windows deliberately. Connect memory, files, databases, prior runs, and custom compaction instead of accepting a black-box prompt stack.
02CONTROL
Draw the loop, not just the prompt.
Compose agents, routers, tools, human checkpoints, retries, and subgraphs on one canvas. Every edge is visible and every node is yours.
03EXECUTION
Run with real boundaries.
Give each tool the exact credentials and permissions it needs. Sandbox execution locally or in the cloud and keep every action attributable.
A COMPLETE AGENT WORKBENCH
Design it.
Run it.
Understand it.
Canvas
Design agents, tools, context, and orchestration on an infinite plane.
↗Runtime
Execute durable turns locally, on your infrastructure, or in managed compute.
↗Ledger
Inspect every model call, tool action, token, cost, permission, and diff.
↗Interface
Use the same project through the GUI, CLI, MCP, API, or a published app.
↗TURNS04 / 08
CONTEXT18,204
COST$0.42
ELAPSED05.1s
00:00.000turn.startedrequest / resolve vendor exception●00:00.018context.mountedpolicy · history · 4 files · 18.2k tokens✓00:01.241agent.decisiondelegate evidence collection → researcher✓00:01.302tool.permissionbrowser.read · vendor portal · allowed✓00:04.886tool.result3 records attached · 1 discrepancy found✓00:05.102human.approvalexternal update requires operator review◫AGENTS.mdAutomatic learning diff
+ Always attach the vendor's original response before proposing a correction.EXECUTION LEDGER
Every execution.
Nothing hidden.
Replay the exact context, model calls, tool inputs, outputs, approvals, token usage, cost, and file changes behind each run.
- 01Compare any two runs
- 02Trace failures to the exact turn
- 03Review what the agent learned
- 04Resume locally or in the cloud
THE OPEN LAYER
Own the whole
agent stack.
The editor, runtime, project format, node SDK, execution ledger, and local runner are open. Build your own nodes, host the entire system yourself, and keep your data where it belongs.
OPEN SOURCEEditor + runtimeFREE
BRING YOUR OWNModels + subscriptionsYOURS
OPTIONALManaged compute + publishCLOUD
# 100% open source · self-host or bring your own modelsFIRST PUBLIC RELEASE COMING SOONWHO BUILDS WITH IT
For teams going
beyond the demo.
When the agent needs domain context, real credentials, durable execution, and a human operating model, the framework is no longer enough.
Forward deployed engineers
Turn what you learn in the field into reusable agent systems without hiding the operational details.
AI-native companies
Give the entire team one place to design, run, inspect, and improve the agents behind the product.
Vertical agent startups
Start with your domain logic instead of rebuilding context, tools, auth, logs, and deployment from zero.
THE NEW WAY TO BUILD COMPUTERS
The next computer
won't be programmed
one prompt at a time.
Build the system behind the agent.
Open the plane ↑