PR-Agent is an open-source code review project. It hit GitHub's daily trending list with 122 new stars in a day, reaching 13,850 total S¹. Its own description carries a pointed disclaimer: "This project is not the Qodo free tier" [S1, S2].

My read: The gap between the 13,850 stars on the trending snapshot S¹ and the 13,550 the organisation's GitHub page recorded at capture P³ is roughly three days of growth at the current rate, which tells me the spike is real and recent. I'm skeptical of the "~30 seconds, low cost" claim per tool S¹ because the README gives no benchmark, no model breakdown, and no error bars. It could hold for a small diff on a fast model and fall apart on a 500-line change. What I find genuinely useful is the identity line the repo is drawing between a community-maintained codebase and a commercial product that shares its DNA.

A legacy project drawing a boundary

PR-Agent is an AI-powered code review agent that the maintainers describe as a community-maintained legacy project of Qodo [S1, S2]. The distinction matters because Qodo, formerly CodiumAI, also sells a primary commercial code review product that the README calls "feature-rich" and "context-aware" [S1, S2]. The open-source project is the earlier codebase; the commercial product is the newer, separate offering.

The relationship is not hostile. Qodo's free version for open-source projects appears as a gold sponsor of PR-Agent [S1, S2]. The Docker image history tells the same story of a gradual split: releases from version 0.34.2 onward live under the pragent/pr-agent namespace on Docker Hub, while older releases up to v0.31 remain frozen at the legacy codiumai/pr-agent namespace [S1, S2].

The latest release, v0.49.0, shipped on 25 October with a change that requires GitHub polling deployments to set an explicit allowlist of repositories P⁵. The GitHub organisation behind the repo was created in late February 2026 and holds two public repositories P³.

What the tools actually do

The project ships three core commands: /review, /improve, and /ask [S1, S2]. Each runs a single LLM call, which the maintainers say takes about 30 seconds at low cost [S1, S2]. No one outside the project has independently verified those figures.

It works across GitHub, GitLab, BitBucket, Azure DevOps, and Gitea [S1, S2]. Deployment options span CLI, GitHub Actions, Docker, self-hosted instances, and webhooks [S1, S2]. On the model side, it supports OpenAI GPT, Anthropic Claude, Google Gemini, DeepSeek, Mistral, and anything reachable through LiteLLM [S1, S2], the same kind of multi-provider routing layer as Otari's open-source LLM gateway, covered last month. For a team running a private GitLab instance that wants automated review comments without sending code to a third-party SaaS, the self-hosted Docker path is the one to check first.

Point it at a cheap model like DeepSeek, run /review on a real pull request, and measure the actual latency and token cost against the 30-second claim. That is the only way to know whether the README's performance figures hold for your codebase.

The broader pattern matches what Cline's open-source coding agent showed when it hit 70,000 stars last month: developers are reaching for tools they can run on their own infrastructure with their own model keys, even when a commercial version exists upstream.

The v0.49.0 release notes P⁵ show active work on polling configuration and Jira integration. The next release is the checkpoint for whether the allowlist requirement stabilises.

Sources: S1 — The-PR-Agent/pr-agent: 🚀 PR Agent: The Original Open-Source PR Reviewe · S2 — The-PR-Agent/pr-agent: 🚀 PR Agent: The Original Open-Source PR Reviewe · P3 — The PR Agent · P4 — Showing how to use huggingface models with PR-Agent · P5 — Releases · The-PR-Agent/pr-agent · GitHub · P6 — adding huggingface inference support + litellm debugger by krrishdhola

Written from 6 sourced items, 4 of them primary.