PUPPETMASTER.GG

MULTI-AGENT ORCHESTRATION · INFERENCE · RESEARCH

PULL THESTRINGS.

Puppetmaster runs agent workers as durable jobs: leased, checked and written to state that outlives the chat. Around it, a desktop harness, inference kernels for 12 GB cards and research published with its limits.

Browse everything

puppetmaster-ai 1.32.5 on PyPI467 stars on GitHub

PUPPETMASTER / JOB CONSOLEREAL RUNcomplete
job_338f29ad552d · analysis · read-only · state.sqlite3
completeconflict-auditorgpt-6.1-sol · needs 85
completeexploregpt-6.1-sol · needs 73
completetest-coverage-reviewergpt-6.1-sol · needs 75
11findings
2risks
6verifications
3decisions
3routing

+54sfindingconflict-auditor[High] The recursive-force delete gate misses ordinary equivalent spellings…

+62sfindingexplore[High] Full-auto recursive-delete gate misses normal equivalent flag spellings…

+67sfindingtest-coverage-reviewer[High] The main classifier regression matrix leaves device-redirect…

+67sjobstitching · summary written · complete

Event table of a real three-worker audit, time-compressed · 67 s · 25 artifacts
BUILT BY CARY PALMERLEASED WORKERS ON CLAUDE CODE · CODEX · CURSOR · ANTIGRAVITY · HERMES · ANY PROVIDER KEY21 ENTRIES INDEXED ↓
10,000live worker processes held on one rented RTX 4090Writeup → 262,144token context served on a 12 GB RTX 4070Mirai S ↗ ~47%lower cost than the frontier-only arm on SWE-bench LiteStudy ↗ 95%of model calls avoided when 1 in 20 turns is novelDurable state ↗
THE SUITE

One ecosystem.
Every string connected.

Start with the runtime. Pick the surface you like to work in. Keep your context wherever you build.

THE FOUNDATION

Puppetmaster

Give your agents a backbone.

  • Leases, not hope.Workers renew a lease on every heartbeat. A dead worker forfeits its task and the task is recoverable.
  • Typed artifacts.Findings, risks and verifications carry confidence, citations and a content hash.
  • Flow graphs.Write one graph. Items repair alone and the pilot wakes only when it matters.
  • Zero-token recall.Pull a finished job into a later session without another model call.
Meet Puppetmaster ↗PyPI ↗
QUICKSTART · bash

# Give the swarm somewhere to live.

$pipx install puppetmaster-ai

$puppetmaster setup

# Then, from any MCP host:

puppetmaster_flow run → workers lease → artifacts persist

THE HARNESS

Marionette

A desktop coding harness with the Puppetmaster kernel inside. Any model pilots. Leased worker swarms, cost-aware routing and a receipt for every run. The console above is this run.

Open Marionette ↗ MEMORY

Portable LLM Wiki

Your context, in your git. Queryable by any LLM over HTTP and MCP.

portablellm.wiki ↗ THE NATIVE INTERFACE

Automaton

Named automata in native GPUI. Direct your staff while jobs live beyond the conversation.

Discover Automaton ↗ REMOTE CONTROL

Discord OS

Discord is the screen. Your computer does the work. Your phone is the remote.

Open Discord OS ↗
KERNELS & INFERENCE

Big models.
Twelve gigabytes.

Frontier-length context on a consumer card. Same weights as published; every number is a measured run.

MIRAI S ON 12 GB

The full 262,144-token window.

Tiered q8_0 cache, MTP drafting at every depth and harness-proofing, served on one RTX 4070.

Model and Mirai S codec by Mirai Labs, on Qwen3.8-27B by Qwen. GGUF and llama.cpp port by alesha-pro.

Context
262k
Decode
76.6tok/s
HumanEval
158/164
Receipts ↗Engine ↗
bonsai.puppetmaster.gg

BONSAI

Ten recorded browser workflows driven by Bonsai 2 on one RTX 4070: CRM, email, issue tracking. Original-speed footage and run evidence.

Bonsai 2 model by PrismML.

Films
10
Decode
54–97tok/s
Tool calls
9/9
Watch the films ↗ KERNEL

Bonsai ADA surgery

Bonsai 2 27B at 262k on 12 GB. Same ternary weights, 2x prefill, Windows bundle.

Model by PrismML; MTP head by ProCreations.

 LAB NOTEBOOK

Scrap VRAM lending

MoE experts lent across scrap GPUs on a home LAN. Prior art cited first.
RESEARCH

Measured, then published.
Limits included.

Studies that test the thesis, including the results that cut against it.

strongorc.puppetmaster.gg · interactive benchmark

STRONGORC

Which models keep delegated work intact? Measured orchestration reliability against total spend, as orchestrator and as worker, with overall, orchestrator and worker Pareto views.

Explore the benchmark ↗
 FROM THE WRITEUPS

Ten thousand live workers

10,000 live OS workers on a rented 4090. Eight decode slots. Not 10,000 concurrent inferences.

Read the writeup →
THE INDEX

Everything,
in one place.

21 projects, papers, benchmarks and notes. Filter here, or press ⌘K on any page.

Suite 5

Kernels 4

Benchmarks 3

Papers 5

Writing 2

Tools 2

The work should survive
the conversation.

Own the machinery.Open source you can read, run and make your own.

Choose your intelligence.Tools designed around agents, not allegiance to one provider.

Keep what matters.Durable jobs, inspectable artifacts and context you can carry forward.