dagraph

v0.1 · MIT & BSL · Self-host ready

Parallel agent
workflows. Hard
budget caps.

dagraph is an open-source DAG orchestrator for Claude agents. Define your workflow in YAML, get parallel execution and a hard USD ceiling per run — without writing scheduler code.

Open dashboard →★ Star on GitHub
pip install dagraph
dagraph run examples/research.yaml --topic "your question"
workflows/research.yaml
nodes:
  - id: research
    type: agent
    model: claude-sonnet-4-6
    prompt: "Research {{topic}}"
  - id: critique
    type: agent
    deps: [research]
    fan_out: 3            # 3 parallel critiques
    prompt: "Find flaws in: {{research.output}}"
  - id: synthesize
    type: agent
    deps: [critique]      # fan-in
    prompt: "Reconcile critiques: {{critique.outputs}}"
budget_usd: 2.00          # hard cap, run halts if exceeded

Parallel waves

Independent nodes run concurrently. Fan-out to N siblings, fan-in to merge — the scheduler handles dependencies.

Hard budget caps

Set a USD ceiling per run. Exceeded? Run halts before the next API call. No runaway loops, no surprise bills.

Self-host or cloud

Run locally with one docker-compose up, or use the hosted dashboard for traces, runs, and team workspaces.