Open Dynamic Workflows
A practical runtime for portable AI-agent workflows across Codex, Claude Code, Gemini, Qwen, Kimi, and other coding-agent CLIs.
Open Dynamic Workflows (ODW) is a TypeScript/Node CLI runtime for portable dynamic workflows: plain JavaScript scripts that fan coding agents out with agent(), parallel(), and pipeline() — outside the host agent’s own context. Claude Code can already run such scripts inside its private runtime; ODW makes the same scripts portable and standalone, so they become artifacts you can version, share, and run on any coding-agent CLI.
Highlights
- Portable across agents — run the same workflow script on Codex, Claude Code, Gemini, Qwen, Kimi, or a custom CLI; switch the underlying agent by switching adapters. ODW only shells out to local CLIs, never calling model APIs directly.
- Claude Code’s dialect, unchanged —
export const metaplus injectedagent/parallel/pipeline/phase/log/args/budgetglobals, with top-levelawaitandreturn. A script written for Claude Code runs here as-is. - Out of context, at scale — the plan lives in code, so intermediate work never pollutes the host’s context and you can fan out dozens of subagents.
- Reliable hand-offs — JSON-Schema structured outputs are validated and retried so multi-stage pipelines compose instead of guessing on free text.
- Background and observable — every run is a detached worker backed by a run directory with
status,logs --follow,result,pause/stop, plus a zero-dependency live web dashboard (odw serve) and a read-only Tauri desktop “observatory”. - Zero runtime dependencies — the engine is async TypeScript (
parallelisPromise.all, no threads); workflow scripts stay plain.jsand ship.d.tsauthoring types for editor autocomplete.
Status
This is the project I am actively building now. The core runtime, CLI, example workflows, and local observability tools are available on main. It is MIT licensed.