Pitroom: cheap workers for an expensive agent
Pitroom
Reading, fixes, tests and reviews go to parallel workers on cheap models; the main agent gets a short answer, a diff, a receipt
- 53 ★ on the repository
- Last commit
- MIT
Broad access
Why this level
- Workers run commands and edit files under your account, and a patch reaches the tree once applied
- Tasks go to the third-party models of the worker providers, including free tiers
Install
- Works in
In your terminal, with SkillFoxx CLI
npx skillfoxx add cli/pitroomDetects the agents on your machine, shows the access level and pins the version.
Other ways to install
Access and checks
How we reviewBroad accessThe tool can change data, run commands or handle secrets.
- Workers run commands and edit files under your account, and a patch reaches the tree once applied
- Tasks go to the third-party models of the worker providers, including free tiers
This is third-party code. Review the repository files before installing.
What it does
Pitroom installs next to Claude Code, Codex, Gemini CLI or Cursor and takes over the heavy part of the work. The main agent states the question, a worker reads the repository, searches and runs commands, and what comes back is an answer, the exact diff and a receipt. Workers are other CLI agents: OpenCode with free models, Codex CLI, Claude Code, Gemini CLI, Qwen Code, Kilo Code, Antigravity CLI and Cursor CLI. Every path:line citation in a worker's answer is checked against the file on disk. Edits go to an isolated copy, so the working tree stays untouched until you apply the patch. The research, implement and review modes are joined by crew, which hands independent tasks to several workers at once, and audit, which asks a second worker to recheck an answer. After each run a receipt prints the worker's token count and an estimate of the savings, and pitroom savings adds them up. The pitroom dash dashboard shows runs on 127.0.0.1, and the same features are available as ten MCP tools through pitroom mcp.
Who it is for. People paying for an expensive model in Claude Code or Codex who want to cut token spend on reading large repositories.
Good fit when
- A repository of hundreds of thousands of lines, and the question is where auto-update lives and which tests cover it
- An audit of several modules for input validation splits into independent tasks and goes out as a crew
- You need a small fix plus a test without touching the working tree until you review the patch
- A worker's answer must be rechecked by a second independent worker before editing
- You want to see how many tokens and how much money runs saved over a week
Not a fit when
- The task is open-ended or design work: workers suit bounded questions, and one open-ended spell check in docs was stopped after 19 minutes
- The project is small and the model reads it in a couple of calls without helpers
- You need confidence that an answer is complete: the audit on the benchmark repository judged some answers partly right
- No CLI worker agent is available: Pitroom does not run without one
Example request
Use pitroom to map how sessions are created and invalidated, then propose a fix.Limitations
You need Node.js 22.13 or newer and at least one worker from the list, OpenCode 2 or later by default. macOS and Linux are the main platforms; on Windows tests run in CI with fake workers and the real workers were not run there. Cursor CLI and Antigravity CLI are beta on Linux and Windows.
Savings figures are approximate: what the main agent would have spent is only estimated. On the benchmark repository the audit judged 5 of 8 answers partly right. Free models share the provider's daily quota, so an answer may not arrive.
How to disable. For Claude Code run claude plugin uninstall pitroom@pitroom, for Codex codex plugin remove pitroom@pitroom, then pitroom uninstall and npm rm -g pitroom.
FAQ
Which models do the workers use?
No model is hardcoded. Pitroom takes the default model of the worker CLI: a free tier, a local MLX or Ollama model, or a cheap paid one. Switch the model in the worker and Pitroom follows.
Can workers damage my working tree?
Runs that change code work in an isolated copy, and the patch is applied as a separate step. A git guard sits on top: a reference-transaction hook refuses every ref update and push URLs go nowhere.
How do I install it in Claude Code?
claude plugin marketplace add ATASTECH/pitroom and claude plugin install pitroom@pitroom install the plugin. To run pitroom yourself, also install the npm package with npm i -g pitroom.
How do I connect Pitroom as an MCP server?
pitroom install --mcp registers the server in the clients it finds, and claude mcp add pitroom -- pitroom mcp works by hand. The ten tools mirror the CLI commands. For clients that take a URL there is pitroom mcp --http on port 7117.
What does the receipt after a run mean?
The receipt shows how many tokens the worker processed, how many it returned to the agent, and the estimated saving against Claude Sonnet. pitroom savings sums all runs. The figures are estimates.
Discussion
No questions or reviews yet.
Related
CLI toolsA terminal environment for several coding agents: each task gets its own branch and worktree and sessions outlive the terminal.
An orchestrator that hands a coding task to CLI workers in separate Git worktrees, reviews and merges the result.
PromptOps keeps Claude Code, Codex and Gemini CLI side by side by project and gives a lead agent MCP to launch helpers
MCP server for a shared agent room: decisions, tasks verified by another agent and a final report, joined by code.
A local control panel for coding agents: it runs Claude Code and Codex, keeps a task queue and merges verified work into main.
A local mesh for coding agents: Claude Code, Codex, OpenCode and Pi exchange questions across projects and machines
Comparison with similar tools
Cards that share the main search query, and how they differ.
| Tool | Official | Repository stars | Main difference |
|---|---|---|---|
| Pitroom: cheap workers for an expensive agent (this card) | no | 53 | |
| Superpowers | no | 316,679 | a development workflow of skills for the main agent, with no worker offloading of reading |
| Claude Code Subagents | no | 1,024 | ready subagents inside Claude Code; no external workers from other CLIs |
README in short
The README covers a quick start per agent, run modes, the dashboard, benchmarks and a worker reference. The benchmarks use the public PI-Desktop repository with 2,349 files: eight questions took 156 seconds, workers read 5.5 million tokens and returned 7,684 to the agent. A separate section lists what the benchmarks do not show.