Parcha Open-Sources AgentRun After Its Greptile Bill Hit $6,451 a Month, Says a Tuned Cheap-Model Reviewer Costs 18 Cents
Parcha has open-sourced AgentRun, an Apache-2.0 platform that wraps existing coding-agent harnesses with run persistence, approvals and spend caps, plus a learning loop meant to cut token spend over time. It supports Claude Code, Codex, OpenCode and pi, and starts locally with npx @parcha/agentrun. It reached the Hacker News front page this week.
The pitch is in the README’s lead case study. According to Parcha, Greptile moved from a flat per-seat price to a dollar per review beyond the first 50 in March. As Parcha’s own agents started opening more pull requests, its monthly review bill rose from $180 in December to $6,451 in October, with $7,018 in November so far. That is about 36x.
What Parcha built instead
Parcha built its own reviewer on particle-1, a cheap model, and tuned it on its own PRs. A review now costs about 18 cents, per the case study. The design is a parent agent that dispatches eight reviewers on a read-only checkout of the PR: five quick ones that read the whole change and three deep ones that dig into risky sections. The parent verifies each finding against the code and posts a single comment-only review. It never approves.
Parcha lists four things that mattered:
- Shell access. With only the diff, the cheap model was wrong about 70% of the time. With read-only bash on the repository, 7%.
- Mixed reviewers. Five quick plus three deep beat eight of the same kind.
- A parent that does not argue findings away. The first version dropped real bugs it could not prove. The final version drops a finding only when the code contradicts it.
- Counting reviews that never posted. A timeout once made a good version look 15 points worse.
The numbers, and why to discount them
On the 30 newest PRs in the AgentRun repository, which Parcha says were held out from tuning (it tuned on 70 merged PRs), the reviewer caught 42% of real bugs against Greptile’s 26%. On the 8 bugs classed as serious it caught 5; Greptile caught 1.
Several things limit what that proves:
- It is Parcha’s own repository and Parcha’s own benchmark. The answer key is 78 bugs, built from every real bug that any reviewer found, so it is a pooled-recall measure rather than ground truth.
- The grader was Claude Opus 5.5, blind to which reviewer wrote each finding.
- Eight serious bugs is a tiny sample. 5 versus 1 is suggestive, not statistical.
- Nothing here measures false positives against Greptile’s configuration. Parcha’s reviewer was tuned on this team’s own PRs, and the write-up does not say how much configuration Greptile received.
Still, the cost mechanics are credible and match what other vendors have been reporting. CodeRabbit said earlier this month that its best single reviewer model catches only about half of known bugs, and it uses much cheaper subagents under a stronger lead. Parcha’s version of the same architecture is a cheap fleet under a verifying parent.
The second case study: learn once, hand off
The other example in the README is merchant classification. Parcha let Opus work out how to classify merchants, then handed what it learned to a cheaper model. Cost fell from $0.89 to $0.12 per merchant, roughly 7x, for the same job. In AgentRun terms, after each run the agent keeps a note, a skill or a workflow that makes the next run cheaper.
That loop is the actual product. The control plane (approvals, spend caps, a single console, resumable runs that survive a server kill) is table stakes. The claim that matters is that a frontier model can be the teacher and a cheap one the executor, and that the platform can manage the handoff. Parcha provides two data points, both its own, and no public benchmark. Anyone adopting it should run the template against their own repository before believing the 18-cent figure.
AgentRun needs Node 24.22.3 or newer on macOS or Linux, and one model key. A single OpenRouter key gives every harness a model, per the README.