Open Dynamic Workflows
Orchestrate coding agents in detached background runs with portable JavaScript workflows.
- Source repo
- xz1220/open-dynamic-workflows
- Stars
- ★ 115
- Last updated
- 1mo ago
- License
- MIT
- Primary language
- TypeScript
- FA score
- 69/100 · Some gaps
At a glance
- How it runs
- Works with
- Universal · cross-platformCodex · Claude Code
- Cost
- Free software; you pay for model usage
- Setup effort
- Low · running in minutes
- You'll need
- Typical use
- Software engineers comparing implementation approaches can use
parallel()to draft alternatives and have an agent synthesize them. - Not a fit if
- Users who need only a single CLI and no workflow orchestration
- Website
- xz1220.github.io/open-dynamic-workflows ↗ set on GitHub
- Maintenance
- Commits in the last 6 monthslast commit 2026-10-19latest release v0.4.0 (2026-06-11)
- Source review
- 69/100 · Some gaps
What does this agent do, and when should you use it?
Open Dynamic Workflows (ODW) is a TypeScript/Node CLI runtime for dynamic workflows written in JavaScript. Scripts use agent(), parallel(), and pipeline() to dispatch installed coding-agent CLIs, including Codex, Claude Code, Gemini, Qwen, Kimi, Oh My Pi, Kilo Code, OpenCode, Cursor, or a custom CLI. Each run executes in a separate background worker and stores its status, logs, results, and changes in a run directory. Users manage runs with the odw CLI or inspect live DAGs, logs, and results in a zero-dependency web dashboard. ODW offers prebuilt binary installation; the agent CLIs it calls must be installed separately.
Users write a workflow script or create one with odw new; the script exports meta and uses injected globals such as agent, parallel, pipeline, phase, log, args, and budget. ODW starts configured agent CLIs in the background and supports concurrent tasks, staged pipelines, structured-output validation and retries, exec() commands, human approvals through checkpoint(), and nested workflow() calls. During a run it stores events, logs, agent results, and diffs from isolated worktrees in the run directory. Users can inspect or apply results with odw status, odw logs, odw result, odw diff, and odw apply, or watch the run in the odw serve dashboard.
- Software engineers comparing implementation approaches can use
parallel()to draft alternatives and have an agent synthesize them. - Teams seeking cross-CLI review can have one adapter implement a change and another adversarially review it.
- Maintainers splitting a change into separate tasks can use isolated worktrees to run tasks in parallel, then review and apply the resulting diffs.
- Users coordinating research or engineering stages can pass items through
pipeline()and require structured JSON outputs. - Teams that need a human gate before execution can pause at
checkpoint()to approve, edit the payload, or reject it. - Developers tracking long-running agent jobs can run
odw serveto view live DAGs, timelines, logs, and results in a browser.
Key setup facts
Pulled from the install and usage notes below. The project README has the current values.
- Install / run
curl -fsSL https://raw.githubusercontent.com/xz1220/open-dynamic-workflows/main/scripts/install.sh | sh
- Runtime
- Node.js 20+
How do you install or deploy this agent?
The simplest installation runs the installer script, which downloads a prebuilt odw binary for the platform and installs the workflow skill. Node.js is not required for the binary install, but you must separately install at least one agent CLI for ODW to call, such as claude or codex. For manual installation, download the archive for your OS and architecture from Releases, verify it against that release's SHA256SUMS, place the binary on your PATH, and copy the skill from the same release. Source development or the not-yet-published npm package requires Node.js ≥20; Git 2.17 or newer is needed for worktrees.
curl -fsSL https://raw.githubusercontent.com/xz1220/open-dynamic-workflows/main/scripts/install.sh | sh
odw --versionHow do you use this agent?
Check the agent CLI setup first; if you have several CLIs, use odw init to select a default, then run odw doctor to inspect the environment. Create and test a template workflow, then use a dry run to inspect its structure before a real run. Add --wait to wait for completion and print the result. Run odw serve --open to open the live dashboard. Model or service charges for real agent calls depend on the selected CLI; the documented dry run uses stubs and does not call a model.
odw init
odw doctor
odw new hello
odw test hello
odw run hello --dry
odw run hello --wait
odw serve --openWhat are this agent's strengths and limitations?
- The same workflow scripts can run through adapters for multiple coding-agent CLIs, including mixed adapters across phases.
- Detached workers provide run status, followable logs, pause, stop, resume, and live DAG observation in a browser.
- Isolated worktrees, diff review, and apply commands let users inspect agent changes before landing them in the working tree.
- JSON Schema validation, retries, and human checkpoints support multi-stage workflows with explicit verification and approval.
- The runtime has zero dependencies, and the prebuilt binary install does not require users to install Node.js.
- At least one supported coding-agent CLI must be installed and configured separately; ODW does not provide a model.
- Real agent calls may incur model-service charges, which are not covered by ODW.
- Some features are ODW extensions, so scripts need compatibility checks before moving to Claude Code's native Workflow runtime or another host.
- Nested
workflow()calls are limited to one level; isolated worktrees start from committed HEAD and exclude uncommitted changes. - The prebuilt binary is about 110 MB on disk (about 35 MB compressed for download).
How does this agent compare with similar options?
Compared with Claude Code's built-in Workflow tool, ODW runs similar workflow scripts as a standalone CLI, supports multiple coding-agent CLIs, and provides detached run management and browser observation outside the host session. The README also notes that Claude Code's native Workflow tool ignores several ODW extension options.
Key facts side by side with the most closely related agents.
| Agent | Source review | Form / cost | Stars | Updated | Language | Full support on |
|---|---|---|---|---|---|---|
| Open Dynamic Workflows This agent | 69 · Some gaps | CLIFree + model costs | ★ 115 | 1mo ago | TypeScript | Codex · Claude Code |
| Pi Dynamic Workflows | 72 · Some gaps | Agent plugin / skillFree + model costs | ★ 597 | 8d ago | TypeScript | — |
| OpenTeams | 55 · Major gaps | Desktop appFree + model costs | ★ 593 | 11d ago | TypeScript | Codex · Claude Code |
| Maestro-Flow | 52 · Major gaps | CLIFree + model costs | ★ 577 | 26d ago | TypeScript | Codex · Claude Code · Claude API |
How does FollowAgents rate this agent?
Why each dimension lost points
Permission tiers, read mode, worktree isolation, human checkpoints, command timeouts and output caps, plus installer checks and rollback are described or supported by code tests, supporting ordinary controls and earning mostly 2s. Deductions: `exec` can run shell commands and workflows launch external CLIs; the sample does not fully establish credential handling, end-to-end data flows, or every external-effect boundary. The repository names an author, but publisher identity and an accountable update path are not independently established, so source attribution earns 1.
The TypeScript structure, CI matrix, conformance suite, fake-CLI tests, and real Docker/SSH CI jobs provide mutually supportive reliability evidence; errors and validation failures are also described as observable. Deductions: this static packet cannot establish that the jobs actually pass, and README and skill coverage differ, so complex behavior cannot receive full marks based on claims alone.
The README and skill cover multiple coding-agent CLIs, parallel and pipeline patterns, and varied workflow scenarios. The skill also says not to use the tool for work that fits in one call, making its audience and trigger reasonably clear. Deductions: conversational triggers remain broad, and although dialect compatibility, extension markers, and permission options are documented, the sample does not give complete platform-by-platform capability boundaries.
The repository gives substantial directory, installation, Node-version, binary/checksum, example, template, limitation, and MIT-license information. Deductions: the README sample is truncated, limiting assessment of its information architecture; naming uses both `odw` and Open Dynamic Workflows; version metadata and release automation are visible, but changelog contents and a clear maintenance owner or support commitment are not, so maintenance responsibility earns 1.
The product combines composable workflow scripts, concurrent dispatch, background runs, DAG observability, structured outputs, and human checkpoints for a clear multi-stage agent-work value. Dry runs and templates also lower trial costs. Deductions: static evidence cannot establish actual output quality or operating costs at scale, and orchestrating multiple CLIs adds setup burden.
Some README claims have corresponding evidence in the skill, exported API, test excerpts, and CI configuration; test comments also specify where fake CLIs are used. Deductions: the packet omits full implementations, the complete test suite, and changelog contents, preventing item-by-item verification; performance and validation claims remain documentary assertions rather than observed results.
- Workflows can run shell commands and dispatch real CLIs. Review scripts, permissions, and the target workspace before running, especially `exec`, network access, and agents that may modify files.
- The install instructions offer a `curl | sh` path; users who prefer a reviewable install can follow the manual-download steps and check the release SHA256SUMS.
- This assessment uses only the supplied static sample. No code was run, and publisher identity, binaries, and CI results were not independently verified.
Recent releases
Latest entries from the project's GitHub Releases. Notes are quoted from upstream as written, not rewritten or tested by FollowAgents.
-
v0.3.0 v0.3.0 — Claude Code's own workflow jobs in the Jobs tab
Release notes excerpt
The Jobs tab now also surfaces Claude Code's own dynamic-workflow runs — both finished history and live in-flight jobs — read-only, merged alongside ODW's own runs. One observatory watches every dynamic workflow on your machine. - A RunSource provider seam merges ODW's RunStore and a new read-only ClaudeRunSource (~/.claude/projects) behind the unchanged /api/runs. - Terminal journals → full DAG; live runs detected from the in-flight journal. Control on a Claude run is refused (409 — observe only). …View on GitHub ↗ -
v0.2.4 v0.2.4 — Claude Code workflows in the dashboard
Release notes excerpt
What's new in v0.2.4 fix(dashboard): Claude Code workflows now show, grouped by provider The Workspace used to drop any workflow that was *shadowed* by a higher-precedence same-named one. So if a Claude workflow shared a name with an ODW workflow (e.g. both deep-research), the Claude one vanished — the desktop app and odw serve dashboard could show zero Claude workflows even though ~/.claude/workflows was being read. …View on GitHub ↗
FAQ
Does ODW itself cost money?
Which agents do I need to install?
claude or codex. You can select among multiple adapters or mix them by phase; ODW does not install those CLIs.Will parallel agent changes directly modify my working tree?
isolation: "worktree" on agent() calls to run agents in isolated worktrees. ODW saves their diffs for review with odw diff and application with odw apply. The README says worktrees start from committed HEAD and exclude uncommitted changes.What happens if an agent does not complete its task?
expect to declare checks such as a command, file changes, or reply pattern. ODW retries with the missed checks included in the prompt, then reports failure if they remain unmet. Inspect the run directory through odw status, odw logs, and odw agents.Can ODW completely replace Claude Code's Workflow runtime?
validate(), exec(), and checkpoint() as ODW extensions and limits nested workflows to one level, so check the compatibility list when migrating scripts.Is Open Dynamic Workflows open source?
Is Open Dynamic Workflows safe to use?
Is Open Dynamic Workflows still maintained?
Who makes Open Dynamic Workflows?
What is Open Dynamic Workflows's official website?
Thanks — we use this to improve the page.