Updated on
Sourcery: reviews and analysis
The cheapest pull request reviewer in the segment, with an editor extension included.
Affiliate link · no extra cost to you
Our verdict
Sourcery comes from the world of automatic Python refactoring and today reviews pull requests in several languages for twelve dollars a seat, the lowest rate in the segment on annual billing. In exchange it looks less deep than its expensive rivals: the most detailed public test of it, run by a Cockpit developer on real changes, ended with the team switching it off. It pays off as a cheap safety net, not as the main reviewer.
Best for: Teams on a tight budget who want an automatic first pass with humans genuinely reviewing behind it.
What the internet says
Sourcery is the cheap entry point of the segment, twelve dollars per seat per month on annual billing, and the only one with a long past outside pull request review: it comes from refactoring Python in the editor and its Visual Studio Code extension is past a million installs rated 4.3 out of 5. The most detailed public test of it, run in May 2025 by a Cockpit developer over four weeks of real pull requests, called close to half of its comments noise and ended with the team switching it off.
What the web repeats in favour
- It is the lowest rate in the segment by a distance, and on public repositories it costs nothing with no time limit
- It does not live only in the pull request: the Visual Studio Code extension is past a million installs rated 4.3 out of 5 and the JetBrains plugin is around 158,000 downloads, so review also arrives before the change is opened
- The change summary and diagram are what people thank it for most: they place a big pull request before you read a line
- Under the model there is an engine of its own that predates LLMs: 212 refactoring rules maintained since its Python tool days
What the web repeats against
- The most detailed public test on real code, by the Cockpit team in May 2025, concluded that close to half of the reviews were noise and a quarter bikeshedding, and they ended up switching it off. On four deliberately planted bugs it found two, against three for GitHub Copilot's review
- The big figure on the pricing page is the annualised one: paying by the month costs fifteen dollars and that only shows when you hover an asterisk
- The entry plan has a per-pull-request cap and a rolling seven-day budget measured in diff characters, and when they run out the review simply does not happen
- The public repository carries over a hundred open issues, with repeated complaints of PyCharm freezes and high processor usage
- The models writing the reviews are third-party, OpenAI and Anthropic per its own documentation, and only the Team plan lets you bring your own
Sweep sources: Official pricing · Docs · App stores · Communities · GitHub · Docs
Pros / Cons
Pros
- Twelve dollars per seat per month on annual billing, the cheapest in the segment
- Free with no time limit on public repositories
- Editor extension with over a million installs on Visual Studio Code
Cons
- The most detailed public test on real code called it noisy and ended with it switched off
- The big figure on the site is the annual one: paying by the month costs fifteen
- The entry plan has a change-size cap and a weekly budget
TLDR: Sourcery reviews every pull request on GitHub or GitLab and also lives inside the editor, where it was born seven years ago as an automatic Python refactoring tool. It carries the lowest rate in the segment, from $12 per seat per month, and costs nothing on public repositories. The depth is what it is: the only serious public test on real code called it noisy, and the team that ran it switched it off.
What Sourcery is and how it works
Sourcery did not start as a pull request reviewer. It started in 2019 as a tool that read Python in the editor and proposed mechanical rewrites, the sort that turn a loop into a comprehension or split a function doing three jobs. That engine is still there and still maintained, with 212 refactoring rules of its own that depend on no language model, and it is the part of the product no competitor in this segment has.
On top of that base the company built what it sells today: an automatic review triggered when a pull request opens on GitHub or GitLab, including the versions organisations run on their own servers. Each review brings three things, a summary of the change, a diagram placing what it touches and where, and line-by-line comments. Language coverage is no longer what its origins suggest: the documentation states that it reviews any language GitHub recognises as one, skipping generated files and dependency lockfiles.
It helps to know where the judgement comes from. The documentation itself explains that Sourcery sends the relevant sections of the diff and the surrounding context to two model providers, OpenAI and Anthropic, under a zero-retention agreement with one and a thirty-day ceiling with the other, and with none of that code used for training. Bringing your own model is a higher-plan feature. That is more transparency than the segment average offers, and also a reminder of where the in-house engineering stops.
What it does not cover marks its territory as much as what it does. There is no Bitbucket and no Azure DevOps here, so a slice of the corporate market is ruled out before price enters the conversation.
What using it is like day to day
Setup is quick and the first review shows up on its own with the next pull request. Anyone working in Python or JavaScript will also meet the other half of the product, the editor half: the Visual Studio Code extension is past one million two hundred thousand installs with 4.3 out of 5 across 43 ratings, and the JetBrains plugin sits around 158,000 downloads. Both shipped updates in July 2026, so the product is alive and maintained.
On review quality there is one public test worth more than twenty blog comparisons, and it does not flatter the tool. In May 2025 Martin Pitt, of the Cockpit team, published what happened when he tried Sourcery and GitHub Copilot review on their projects. He first built an example with four deliberately planted bugs: Copilot found three, Sourcery stopped at two, and classic static analysis saw only one. So far, a reasonable draw.
The trouble started once it ran against real work. Going back over every pull request from the previous four weeks of cockpit and cockpit-podman, his conclusion was that about half of the AI reviews were noise and a quarter bikeshedding, and that of the rest, half were small useful hints and half were outright wrong. The sharpest complaint is not about the model but about the product: when it has nothing to say it writes a comment anyway and fires a notification, and there was no configuration knob to turn that off. The team ended up dropping it and now asks for a Copilot review only when a human thinks it will help.
That verdict needs context before it becomes a sentence. It is one team, with mature projects and high standards, and the author declares his own scepticism about AI. But it is the most serious field evidence that exists on this tool, it is dated and you can read all of it, which is more than can be said for the numbers the vendors in this segment publish.
The public repository carries over a hundred open issues, and there you see the other side of daily use. Freezes of the JetBrains plugin come back repeatedly, with one case opened in June 2026 and a similar one from 2024, along with high processor usage attributed to the command line tool. None of that touches review inside the pull request, but it matters to anyone buying the product for its editor half.
Pricing and plans
The entry paid plan starts at $12 per seat per month, and that figure carries fine print the site tucks into an asterisk. Hover it and the full note appears: $12 billed annually, $15 paying month to month. The documentation puts it more plainly in a table, with Pro at $15 a month or $12 annual, and Team at $30 or $24. We publish the figure the page shows, but if you are paying monthly, yours is the other one.
Public repositories pay nothing, with no time limit. For private ones there is a fourteen-day trial with no card, during which everyone in the organisation gets access.
The part almost nobody checks before signing is the usage limits, and here they are declared with uncommon honesty. Sourcery meters consumption in diff characters, meaning the size of the changed code it reads, and applies two caps: one per pull request and a rolling seven-day one. On the entry paid plan that is 300,000 characters per pull request and 1,500,000 across the week, multiplied by the number of seats, with a re-review of the same change costing a tenth. When the budget runs out the review simply does not happen, though it never blocks a merge. A team working in small changes will never notice this. A team moving large refactors will find it is the plan’s real ceiling.
Who it is for (and who it is not for)
Sourcery makes sense when budget rules and automated review is understood as a first pass rather than the main safety net. A small team that wants change summaries, a few mechanical warnings and editor help at the lowest price in the segment has it here. Open source projects are outright invited, since they pay nothing.
It also fits shops carrying a lot of legacy Python, for a reason no comparison table mentions: those 212 deterministic refactoring rules are still there, they do not hallucinate and they work in the editor before a pull request exists.
It is not for anyone expecting the machine to catch what an experienced reviewer misses. The only public field test says that on mature projects it adds little and interrupts plenty, and that kind of team is better off paying more for Greptile or sticking with whatever review already comes bundled. Nor is it for code hosted on Bitbucket or Azure DevOps, or for governing complex house rules, which is Qodo territory. The full ranking sits in the best AI code review tools, inside the best AI coding tools.
Alternatives to Sourcery
The most direct comparison is CodeRabbit, which costs well over double but covers more platforms and has far more checkable user experience behind it. If you would rather pay for what you consume than per person, Qodo sells shared credit pools and mines rules from the team’s history. And if the reason to automate review is security, DeepSource puts deterministic static analysis ahead of the model, which is exactly what Sourcery does only in its old refactoring half.
Frequently Asked Questions
What does Sourcery really cost?
From $12 per seat per month if you pay a year up front. Month to month it is $15, and that only shows when you hover the asterisk on the pricing page. Public repositories pay nothing.
Is it useful beyond Python?
Yes. It was born as a Python refactoring tool and that engine is still alive, but pull request review claims to cover any language GitHub recognises as one.
Can a large change go unreviewed?
It can. The entry plan caps a single pull request at 300,000 diff characters and the rolling week at 1,500,000, multiplied by seats. Once that is spent the review is skipped, though it never prevents a merge.
Where does my code go?
To OpenAI and Anthropic, per its documentation, with zero retention agreed with one and a thirty-day ceiling with the other, and with no training use. Bringing your own model or self-hosting requires the higher plan.
Is it better than GitHub Copilot review?
In the only detailed public test available, no: on four planted bugs Copilot found three and Sourcery two, and on real pull requests the team kept Copilot on request and switched Sourcery off. It is a small sample, but it is the sample there is.