PR Agent vs Qodo: one LLM call per slash command, prompts in TOML
🚀 PR Agent: The Original Open-Source PR Reviewer. This project is not the Qodo free tier.
At a glance
- What is it?
- PR Agent is the MIT-licensed, self-hosted PR reviewer behind slash commands such as /review and /describe, and it is explicitly not the same product as Qodo's hosted code review offering. This is what the architecture, the install paths, the dependency pins and the currently disabled /help_docs command actually tell you.
- Who is it for?
- PR Agent fits a team that already holds a model key and wants to own its review prompt in pr_agent/settings/configuration.toml, and it does not fit a team that needs vendor support for the reviewer itself or a documented retry path when the provider times out.
- Can I use it commercially?
- Yes. MIT is a permissive licence: you can use, modify and sell software built on it, as long as you keep its copyright and licence notices.
- Is it still maintained?
- Yes. The repository received new commits within the last day.
- What is it written in?
- Mainly Python, according to GitHub's language statistics.
Answers come from the project's GitHub data, last synced on November 15, 2026, and from our analysis. They are not legal advice.
Star history
- The-PR-Agent/pr-agent16,235 ★ · +3,233
| Repository | 2026-09-15 | 2026-11-15 |
|---|---|---|
| The-PR-Agent/pr-agent | 13k | 16.2k |
3,233 Stars in the last 61 days, measured from the daily GitHub snapshot. Compare with other projects
Editorial analysis
One LLM call per command, and what the compression hides
A review bot that answers a slash command is a different shape of tool from one that reads every commit on its own. PR Agent waits to be asked: /describe, /review, /improve and /ask "What does this PR change?" are posted as a pull request comment, and /similar_issue runs on an issue instead to find similar issues already in the repository. The same tools run from a terminal. The design choice worth noticing is the cost shape. The README states that each tool, naming /review, /improve and /ask, uses a single LLM call of about 30 seconds at low cost, so what you pay for a review tracks the number of times you ask rather than the size of the diff. That promise is held up by the PR Compression strategy, which the project points at as the way small and large pull requests both get processed. The compression is where the fine print lives. The README does not publish the token budget for a compressed diff, so a reader cannot tell how much of a large change set reaches the model, and a review that stays silent about one file may be the compression talking rather than the reviewer.
The action file pins nothing, the CLI needs Python 3.12
Two install paths cover most people, and they differ in how much you control. The GitHub Action is the route the README recommends. The workflow below sits at .github/workflows/pr-agent.yml and runs on ubuntu-latest whenever a pull request is opened or synchronized, passing your OpenAI key and the repository token in as secrets:
# .github/workflows/pr-agent.yml
name: PR Agent
on:
pull_request:
types: [opened, synchronize]
jobs:
pr_agent_job:
runs-on: ubuntu-latest
steps:
- name: PR Agent action step
uses: the-pr-agent/pr-agent@main
env:
OPENAI_KEY: ${{ secrets.OPENAI_KEY }}
GITHUB_TOKEN: ${{ secrets.GITHUB_TOKEN }}Look at the uses line. It tracks the default branch, not a tag, so the code reviewing a pull request can change without anything changing in your repository. If you want it held still, the release tags are where you hold it, v0.49.0 being the current one. The CLI route is a pip install and needs Python 3.12 or newer, which is what pyproject.toml declares as requires-python = ">=3.12":
pip install "pr-agent[github]"
export OPENAI_KEY=your_key_here
pr-agent --pr_url https://github.com/owner/repo/pull/123 reviewWithout OPENAI_KEY exported, the command has no provider to call. GitLab, BitBucket and Azure DevOps are wired up by webhook or app instead of a workflow file, and those instructions sit on separate documentation pages rather than in the README.
The Docker namespace split at 0.34.2 breaks pinned image references
Anyone running the reviewer in a container hits a rename the project has not cleaned up for you. Releases 0.34.2 and later are published on Docker Hub under pragent/pr-agent. Everything up to and including v0.31 lives at the old codiumai/pr-agent namespace, which the README calls a frozen archive where no new images are pushed. A self-hosted deployment that kept an image:, docker pull or uses: docker:// reference pointing at codiumai/pr-agent still pulls successfully, and nothing errors. It simply runs a build that stopped at v0.31 while the pip package in the same container moved on. Every release between v0.31 and 0.34.2, and everything since, is missing from that image. The repository ships Dockerfile.github_action_dockerhub and a docker/ directory, so building locally is possible, but the image tag baked into a compose file or a self-hosted runner is the first value to audit, before the version of the package you pip installed next to it.
Half the dependency list is pinned exactly, half takes a range
The dependency block in pyproject.toml is split down the middle, and a comment says the split is deliberate. Several packages are pinned to one version, among them Jinja2==3.1.6, litellm==1.105.1, loguru==0.7.3, PyYAML==6.0.3, tiktoken==0.14.0 and unidiff==1.0.1. Others take a floor and a ceiling, among them dynaconf>=3.3.5,<4, fastapi>=0.141.1,<1, GitPython>=3.1.62,<4 and starlette>=1.6.0,<2. The comment above the list says security-sensitive packages declare >=patched,<next-major ranges instead of == pins so pip consumers can pull security patches on their own cadence, citing issue #2523, with the floor set to the last vetted patched release. One package is pinned twice by interpreter, google-cloud-storage==2.10.0 for python_version < '3.13' and 3.12.0 for python_version >= '3.13', an arrangement added for issue #2480. The consequence for a locked or offline build is that you reproduce both halves by hand, and the exact pins mean a security fix for those packages arrives only when the project moves the number itself.
/help_docs has been switched off since v0.36.1 over a credential leak
One documented command is currently missing. The README marks /help_docs as temporarily disabled since v0.36.1, pending a fix for a credential-exposure issue tracked as #2445. That is the tool that answers questions out of the project's own documentation, and current releases will not run it. Meanwhile setup.py still ships the corpus: its build step walks docs/docs for .md and .mdx files and installs them into the pr_agent/_help_docs package directory, and it raises FileNotFoundError when it finds none, which makes the documentation a build requirement rather than an optional extra. So the corpus travels inside the package while the command that reads it is off. The reason matters more than the missing feature. A credential-exposure issue means that command was reaching something it should not have been, the fix was still pending at v0.49.0, and nothing in the README tells a reader on an older release that it has been resolved. A build with /help_docs enabled is not the one to put back into service, and re-enabling the command on a pinned old tag to get the feature back reopens the reported problem.
No documented retry, no cost ceiling, one provider in the privacy note
The repository covers the working path well and the failure path thinly. At the top level there are action.yaml, the project's own .pr_agent.toml, AGENTS.md, pr_compliance_checklist.yaml, CHANGELOG.md and RELEASE_NOTES.md, and review behaviour lives as JSON prompting inside pr_agent/settings/configuration.toml. What none of that settles is what happens when the provider is the part that breaks. The README does not document a retry count for a timed-out request, a ceiling on tokens or spend per pull request, or a way to make a re-run leave one comment instead of two after a job dies mid-flight. The Data Privacy section is equally narrow: it covers hosting PR Agent with your own OpenAI key and says the traffic is between you and OpenAI, then stops. The routers named elsewhere in the README, including AWS Bedrock, Vertex AI, Databricks, OpenRouter and a local Ollama, are not covered there, so what leaves your network is a question those providers' own documentation has to answer, not this README.
PR Agent and Qodo share a name and a sponsor, not a product
The naming causes more confusion than the code. The README is explicit that this repository is not the Qodo offering for open-source projects, and it calls PR Agent a community-maintained open-source project and a legacy project of Qodo, distinct from Qodo's primary AI code review offering, which is presented as a separate, context-aware product with a free version for open-source projects. Qodo is also the listed gold sponsor. The difference that changes day-to-day work is ownership. Here the reviewer is MIT-licensed, runs on infrastructure you control, and its behaviour is a prompt file you edit in pr_agent/settings/configuration.toml. The model sits underneath that and is chosen through LiteLLM, so the same prompt can point at OpenAI GPT, Anthropic Claude, Google Gemini, DeepSeek, Mistral, or a host such as Azure OpenAI, AWS Bedrock, Vertex AI, Databricks, OpenRouter or a local Ollama. Taking the hosted reviewer gives you a supported product and taking this one gives you the prompt file. A team expecting vendor support for the reviewer itself is choosing between two different things, not between two tiers of one thing.
One command proves the unit layer, not the end-to-end one
The cheapest way to know whether a checkout is sound is its own test suite. The repository carries a tests/ directory and a contributing guide. pyproject.toml reads version 0.49.0, matching the v0.49.0 tag from 2026-10-25, which follows v0.48.0 on 2026-10-16 and v0.47.0 on 2026-10-02, and the default branch was last pushed on 2026-11-02 with the repository not archived. uv.lock, .python-version and .pre-commit-config.yaml are all committed, so the environment is reproducible. What that cadence does not tell you is how fast a change to review behaviour reaches you, because prompts ship as configuration inside the package rather than as a review policy you can version and diff on its own.
Editorial conclusion
PR Agent fits a team that already holds a model key and wants to own its review prompt in pr_agent/settings/configuration.toml, and it does not fit a team that needs vendor support for the reviewer itself or a documented retry path when the provider times out. Before adopting, check three things in order: whether your container still pulls from the frozen codiumai/pr-agent archive, whether the action is pinned to a tag or tracking main, and whether a /help_docs enabled build from an older release is still in your pipeline. The last push was 2026-11-02 and the repository is not archived, so release cadence is not the concern; the open item is the credential fix in issue #2445, still pending in v0.49.0.
Frequently asked questions
What is a PR review agent?
PR Agent is an open-source, AI-powered code review agent that runs slash commands such as /describe, /review, /improve and /ask as pull request comments or from the command line, with each tool using a single LLM call.
What is GitHub PR Review Agent?
For GitHub, the project ships a GitHub Action you add as a workflow file at .github/workflows/pr-agent.yml, using the-pr-agent/pr-agent@main and reading OPENAI_KEY and GITHUB_TOKEN from repository secrets. It triggers on pull requests that are opened or synchronized.
How do you use PR Agent?
Locally you pip install "pr-agent[github]", export OPENAI_KEY, and run pr-agent --pr_url <PR_URL> review, or use pr-agent --issue_url <ISSUE_URL> similar_issue for issue-scoped commands. On a hosting platform the tools are posted as comments on the pull request.
PR Agent vs Qodo: what is the difference?
The README states that this repository is not the Qodo offering for open-source projects, and describes PR Agent as a community-maintained legacy project of Qodo, separate from Qodo's primary AI code review offering. Qodo is also the listed gold sponsor.
What is an alternative to PR Agent?
The README points to Qodo's primary AI code review offering and its free version for open-source projects as a separate hosted product. Inside PR Agent the model layer is swappable through LiteLLM, which reaches OpenAI GPT, Anthropic Claude, Google Gemini, DeepSeek, Mistral, and hosts including a local Ollama.
Official sources
Add this badge to your README
If you maintain this project, the badge below links readers to this analysis and shows its maintenance status from the daily GitHub snapshot. Paste the markdown into your README; add ?metric=license or ?metric=stars to the image URL for a different field.
[](https://hysenlabs.com/projects/the-pr-agent-pr-agent)Star history chart for your README
[](https://hysenlabs.com/projects/the-pr-agent-pr-agent)Was this useful?