Open Dynamic Workflows

The model plans the swarm. A local script runs it. Your chat only sees the answer.

An open-source engine for dynamic, multi-agent workflows where a *script* is the orchestrator, not the model — the pattern behind Claude Code's dynamic workflows and ultracode, for OpenCode, Cursor, OpenAI Codex, Google Antigravity, and VS Code. Bring your own model, run a local daemon, keep everything on your machine.

Quick start · How it works · Topologies · Compare · FAQ

What it is

When you ask one LLM to coordinate fifty agents, it spends its context window keeping track of the other forty-nine. The fix is to have the model write a plan once — a plain JavaScript execute() function — and then step out of the loop while a runtime executes it. The model is the author. The script is the orchestrator. That capability had been locked to one proprietary tool; this is the same idea, MIT-licensed, on your machine.

How it works

The model never babysits the swarm. It writes one execute(context) function and hands it to the daemon, which runs it inside a WASM-isolated QuickJS sandbox where the only things in scope are the workflow primitives — agent, parallel, pipeline, verify, loop, checkpoint. Every agent() call becomes one HTTP request to your model provider, scheduled through a concurrency queue. Your chat window only sees the final answer.

- It survives. State lives in SQLite with write-ahead logging. Kill the daemon mid-run, start it with --resume, and completed agents come back from cache — only unfinished work re-runs. A failed agent is dropped, not fatal; the run keeps going. Node identity is sha1(workflow | phase | role | prompt), so replay is exact.
- It doesn't trust its own agents. A finding isn't a finding until a panel of skeptics has…