graph-execution
Turn a spec into a reviewed, tested pull request.
graph-execution plans a task graph from your PRD, runs coding agents in parallel in their own git worktrees, gates every task, then has a code reviewer and a live tester check the merged result before it opens the PR.
npm install -g graph-execution && graph-execution manager Works with Claude Code, Codex, OpenCode and Qwen Code. MIT licensed. Runs on your machine.
How it works
From spec to pull request, checked at every step.
Most agent tools make code faster. graph-execution makes the result something you can merge.
- Write a spec.
Describe the change in a PRD, or start from Jira tickets.
- Plan the graph.
The spec becomes tasks with dependencies. A task starts only after the tasks it needs have merged.
- Run agents in parallel.
Each task runs in its own git worktree, with the agent you choose.
- Gate every task.
Your build, lint and tests run on each task before it merges.
- Review the whole result.
A code reviewer and a live tester check the merged graph against the spec. A fixer handles blockers and majors. Then you open the PR.
Why graph-execution
Built for output you can trust.
Parallel agents are only useful if what comes back works. Every run ends in review, not in hope.
Verified, not just generated
Every task passes your gate, then a code reviewer and a live tester check the merged result before anything reaches you.
Any agent
Claude Code, Codex, OpenCode or Qwen Code. Pick one per run. No lock-in to a single vendor.
Parallel by default
Independent tasks run at the same time, each in its own git worktree, and merge as they pass.
A web app for the whole loop
graph-manager clones your repos, opens agent terminals, shows every run live and creates the PR.
Multi-repo runs
One run across several repos, with one spec per repo and a joint verification at the end.
Sandboxes
Run agents in their own Docker container instead of with your user's permissions.
Proof
Built by itself
graph-execution is developed with graph-execution. Every feature below started as a spec and shipped as a run.
26 specs shipped
- run-graph skill and the --web run view
- Self-improving REVIEW.md and TESTING.md (improveDocs)
- The loop above the graph (graph-execution auto)
- Live change view (KAN-4)
- Feature test plan (KAN-5)
- Human review round (KAN-15)
- Run cleanup and suite (KAN-21, KAN-43)
- Web UI redesign ("mission control")
- GitHub-style review comments
- Hide the human review step when human review is off
- Count only open findings in review summaries
- Show when an agent was stopped for a loop
- Agent tools: graph-manager MCP tools for agents opened in its terminals
- Bug-hunt fixes GE-01 … GE-13 (Linux/macOS)
- graph-manager design v1
- graph-manager without terminal statuses
- See the spec while it is being created (graph-manager)
- Monorepo and graph-manager
- Landing site at graph-execution.com
- Multi-repo runs: graph-manager projects, one suite across repos, a joint verification
- graph-manager from npm: graph-execution manager
- Toggles for all run options in Start run (graph-manager)
- Sandbox image built by graph-manager
- Sandbox image: Node by default, toolchains on demand
- Sandboxes for runs (graph-manager + engine)
- Add and edit review agents from the UI (graph-manager + engine)
Coming next
Hosted for teams is coming.
Run graph-execution without your own server: sign in, connect your repos and Jira, and let your team turn tickets into reviewed PRs.
Watch the repo on GitHub for the launch