professorpalmer/Puppetmaster

Provider-neutral control plane for durable-state agent swarms: subprocess workers, leases, artifacts, memory, and deterministic stitching.

What it solves

Puppetmaster addresses the lack of durable state and reviewable output when using various AI agent CLIs and provider APIs for complex engineering tasks. It prevents the loss of progress and expensive context window growth by storing agent results in a structured database rather than relying on a single, ever-growing chat transcript.

How it works

It acts as a supervisor and job store that sits between "pilots" (MCP-enabled tools like Cursor Agent, Grok Bot, or Claude Desktop) and "workers" (agent adapters like Claude Code, Codex, or provider APIs).

  1. Routing: It routes tasks to available models or specific worker adapters.
  2. Execution: Independent workers claim tasks and execute them.
  3. Persistence: Instead of sharing a transcript, workers write typed artifacts and evidence to a SQLite database.
  4. Retrieval: A subsequent model can retrieve these stored artifacts to create a stitched summary or continue work without inheriting the previous model's provider KV cache.

Optional integration with CodeGraph can provide structural code indexing to enhance the context provided to workers.

Who it’s for

Developers who perform repository investigations, audits, refactors, and implementations and want a way to inspect, resume, and review the output of multi-step AI engineering workflows.

Highlights

  • Durable State: Uses SQLite to store job results, allowing for inspection and resumption of work.
  • Multi-Adapter Support: Integrates with various tools including Cursor, Claude Code, Codex, and Grok Bot.
  • Cost Efficiency: Demonstrated lower actual spend on SWE-bench Lite through cost routing and durable retries.
  • MCP Integration: Provides MCP tools for pilots to delegate larger tasks to durable workers.
  • Job Dashboard: Includes a live dashboard to monitor and view job progress.

Related

  • Project
    conductor-oss/conductor

    An open-source durable execution platform for microservices and AI agents that turns complex workflow graphs into persisted, inspectable, and recoverable executions.

  • Project
    microsoft/pg_durable

    pg_durable is a PostgreSQL extension that enables fault-tolerant, long-running background workflows directly inside the database, eliminating the need for external job queues or orchestrators.

  • Project
    23blocks-OS/ai-maestro

    AI Maestro is a self-hosted peer-mesh platform that orchestrates AI agents across multiple machines, giving them persistent memory, agent-to-agent messaging, shared task boards, and secrets management. It solves the problem of running many agents in separate terminals with no coordination, and works with existing agents like Claude Code, Codex, and Grok Build.

  • Project
    waynesutton/builder-skills

    builder‑skills provides a set of concise, markdown‑based “skills” that teach AI coding assistants how to follow Convex best‑practice rules and disciplined project‑workflow habits (PRDs, task tracking, safe git). Installable via skills.sh, Claude plugins, or npm, it drops starter docs and lets agents load the appropriate SKILL.md on demand, keeping code, docs, and safety checks in sync.

  • Project
    Cjbuilds/Codex-Orchestration

    Codex Orchestration is a Codex plug‑in that lets you attach other LLMs (Claude Fable 5, Opus 5, OpenRouter models, etc.) to a task and assign them specific roles—Planner, Advisor, Designer, Executor. Codex stays the top‑level orchestrator, iterating plans, reviewing them, optionally designing UX artefacts, and finally executing code. The plug‑in is installed via the Codex marketplace, configured with skill prompts inside a Codex chat, and supports role‑specific effort levels. It aims to improve plan quality, enable parallel work, and reduce premium‑model usage. MIT‑licensed.