Turn Cursor, Claude Code, the OpenAI API, or the Codex CLI into an orchestrator that routes every task to the cheapest model that can handle it, runs workers as independent processes, and stores their output as typed SQLite artifacts so follow-ups cost zero tokens.
💸 Reproduce the live A/B in ~$0.01 of spend—OPENAI_API_KEY=... python -m bench.router_live_ab
. Pinnedgpt-5.5
cost$0.0132; Puppetmaster routed the same task togpt-5.4-nano
for$0.00016(same prompt, equivalent answer). The 35.1% figure is a 6-task mixed-workload dry-run where the routercorrectlykept the frontier model on the 2 hard tasks — full method in[docs/CLAIMS.md].
🔁 Self-healing — a dead provider doesn't kill the swarm (proven live, jobajob_d82715bebc5d
):claude-code
worker hit a real$0 Anthropic balance→ classifiedbilling_or_quota
→ markedFAILED→auto-rerouted to(plan-billed,cursor/gpt-5.5
$0
) → the funded adaptercompleted the task.No silent degraded run.
pipx install puppetmaster-ai # or: pip install puppetmaster-ai
puppetmaster setup # doctor + models init + MCP installers + agent rules, idempotent
That's the whole install. setup
runs every step idempotently, skips any tool that isn't present, and prints what it did. Restart Cursor (or open a fresh Codex / Claude session) and the agent sees 32+ puppetmaster_*
tools plus a rule nudging it to reach for them on multi-file work.
To run benchmarks or hack on it, clone instead — see Contributing. (pipx
keeps the CLI in its own isolated environment, which is the recommended way to install a command-line app.)
New here? Watch the GIF above, run pipx install puppetmaster-ai && puppetmaster setup
, then skim What it does.
| Want to… | Go to |
|---|---|
| Understand the design & what it fixes | |
docs/COMPARISON.mddocs/SECURITY.mddocs/CLAIMS.md· receipts inbench/
docs/FEATURES.mdQuickstart·docs/DAILY_DRIVER.mddocs/README.md·puppetmaster/
·bench/
·examples/
·scripts/
·clients/typescript/
cursor-extension/
Think Redis/Gunicorn for agentic engineering:
Cursor Agent / Claude Code / OpenAI / Codex CLI / shell
|
v
Puppetmaster supervisor ──> task-aware model router (auto-routes by cost)
|
v
independent worker processes ──> SQLite (typed artifacts, events, memory)
|
v
live artifact board ──> stitched summary ──> 0-token follow-up reads
Puppetmaster isn't trying to beat native IDE subagents at every tiny task. It's for the work that gets messy: long repo investigations, conflicting hypotheses, repeated handoffs, flaky memory, and code changes that need evidence, replay, and approval gates. The rationale and failure modes it fixes are in docs/WHY.md.
How it's different: LangGraph, CrewAI, and the Claude Agent SDK are libraries you write code against to build an agent. Puppetmaster sits one layer up — it orchestrates the agent CLIs you already pay for (Cursor, Claude Code, Codex, OpenAI), routes each task to the cheapest sufficient model, keeps the spend inside your subscription, and self-heals when a provider is down. Full side-by-side + "pick X instead if…" in docs/COMPARISON.md.
The whole story in one command — local + shell adapters, nothing to configure:
./scripts/demo.sh # the 60-second tour (clean machine, no keys)
python -m puppetmaster dashboard # live, zero-dependency web board for any job
It routes a task mix by cost, fans out a 6-role swarm as independent processes, reads the stitched summary, then proves follow-up reads cost $0.00. Script + GIF source: scripts/.
Every number is reproducible from a script in bench/. Full detail + caveats:
Cost is fixed on two axes. New work auto-routes to the cheapest sufficient model (35% cheaper on a fixture;98.8% cheaper in a live OpenAI A/B). Follow-ups are SQLite reads, not new agent runs (40 queries, $0.00, 0.5 ms each).** Workers don't share a transcript.They lease tasks and emit typed artifacts**(payload +evidence
+confidence
+sha256
); the stitcher reads JSON, not stdout. Inspect withpuppetmaster artifacts <job_id>
.Graphing is Workers auto-inject task-relevant graph context before the model call; fall back to grep/read without it. (CodeGraph's win, wired in cleanly.docs/CODEGRAPH.md)A dead provider doesn't kill the swarm (v0.9.0+). Billing/quota/auth/missing-CLI failures are markedFAILED
andauto-rerouted to the next funded adapter, preferring plan-billed models. Validated live; surfaced loudly in the summary's Alerts section.
After install, try one of these inside Cursor Agent or Codex:
Use Puppetmaster to run doctor in this repo and summarize what is missing.
Use Puppetmaster to start a cursor swarm for this repo and return the job id immediately.
Problem: users get logged out after refresh and token-refresh tests are flaky.
Constraints: keep the patch focused, preserve public API behavior, run relevant tests.
Do review/plan first. Poll status/logs by job id. Do not edit until you summarize findings and ask for approval.
Or from the shell:
puppetmaster doctor
puppetmaster route "Security audit every endpoint" --role audit # dry-run routing decision
puppetmaster cursor "Review this repo for release blockers" --review --dry-run
puppetmaster claude "Implement the approved change and run focused tests" --permission-mode acceptEdits
puppetmaster show $(puppetmaster last)
More recipes in docs/DAILY_DRIVER.md.
Daily-driver beta. Real runtime contract, automated tests, SQLite default backend, fail-closed jobs, live Cursor Agent MCP, installable Cursor extension, validated full-edit adapters. Credible for supervised local engineering; not yet a hosted multi-user service. Full feature matrix: docs/FEATURES.md.
Pip name: PyPI lists this as puppetmaster-ai because
PEP-503 normalizationcollides
puppetmaster
with an abandoned 2019. The import name, CLI, repo, and brand stay
puppet-master
puppetmaster
. (tracking)
MIT