Free AI Code Review Tools: What You Get for $0 (2026)

Direct answer

Very few AI pull request reviewers are free for private code. Of the 51 tools we track, 9 have a permanent $0 plan that touches private repos, and every one caps you hard - 20 reviews a month, or 1 developer, or 1 summary-only review an hour. Everything else is either free for public repos only, or free because it is open source and you run it yourself. Three distinct categories, three completely different trade-offs. Verified 2026-10-18.

Free AI code review tools to try

Direct answer: cubic Starter (20 PR reviews/mo per GitHub org), Greptile Starter (1 active dev, 50 credits/mo) and Codacy Developer (free forever, IDE-only) are the three private-code free plans that need no application and no card. For public repos, CodeRabbit, Sourcery, cubic and Korbit Max are free. For zero caps, PR-Agent is MIT and Kodus is AGPLv3. Verified 2026-10-18.

Here is the sentence most articles on this topic bury in paragraph nine: free AI code review is abundant for open source and genuinely scarce for private code. Every vendor in this market has worked out that public repos are marketing and private repos are revenue. Once you internalise that, the whole landscape becomes legible.

This article covers permanent $0 tiers only. If you are looking for 14-day trials, card requirements and what happens on day 15, that is a different buying decision and we split it out: AI code review tools with a free trial.

The three kinds of free

Treating “free” as one category is why most roundups are useless. There are three, and they fail in different ways.

Category	What it means	Fails when	Examples
1. Free for public repos only	Full or near-full product, restricted to open source or public repositories	Your code is private, which for most readers it is	CodeRabbit, Sourcery, DeepSource, Korbit Max, cubic, Codacy Team for OSS
2. Free with a hard usage cap	Works on private code, but metered	You grow past 1 developer, or 20 reviews, or 50,000 lines	Greptile Starter, cubic Starter, Codacy Developer, Semgrep Free, SonarQube Cloud Free, Snyk Free, Qlty Cloud Free, Graphite Hobby
3. Free because open source and self-hosted	No licence fee at any size, any repo	You have no one to run it, or no model budget	PR-Agent (MIT), Kodus (AGPLv3), SonarQube Community Build, Qlty CLI, Tabby (Apache 2.0)

The rest of this article takes each category in turn with the actual numbers.

Category 1: free for public and open source repos only

This is the largest and most generous group, and it exists because a public repo is a billboard. The useful distinction nobody draws is install-it-yourself versus apply-first, because an application form is a week of calendar time.

Tool	What the free OSS tier includes	Install yourself or apply?	Platforms	Catch
CodeRabbit	”Free reviews forever for public repositories”, subject to a Fair Usage Policy	Install yourself	GitHub, GitLab, Bitbucket, Azure DevOps	Private repos not included. Most verbose default of the group, so configure before enabling
Sourcery	AI reviews on public repos, PR summaries, diagrams, one-click fixes, custom review rules, auto-approve low-risk PRs, security checks	Install yourself	GitHub, GitLab	Public repos only. Paid Pro is $12/seat/mo annual
DeepSource	Unlimited public repos, 1,000 pull requests reviewed per month	Install yourself	GitHub, GitLab, Bitbucket	AI Review and Autofix are pay-as-you-go on top: $8 per 10,000 lines standard, $15 advanced. Private repos not included
Korbit	Korbit Max free for all open source repos, no published cap	Install yourself	GitHub, GitLab, Bitbucket	Max is Korbit’s top non-enterprise tier, which makes this unusually generous. No published cap means no published guarantee
cubic	Free for public repositories, no review limit stated	Install yourself	GitHub	GitHub only
Codacy	Team plan “free forever for open-source projects”	Install yourself	GitHub, GitLab, Bitbucket	Qualification is per project
Macroscope	Open source free	Install yourself	GitHub	Usage-billed otherwise at $0.05 per KB of diff
Panto	Unlimited pull request reviews for open source repos	Install yourself	GitHub	Panto repositioned around AI mobile QA in 2026 and publishes no per-seat review pricing
Greptile	Free for qualified non-commercial projects with an OSI-approved licence	Apply	GitHub, GitLab	Qualification required. Also 50% off for pre-Series A startups under $2M revenue
Gitar	Full Pro feature set free for OSI-licensed open source	Apply	GitHub, GitLab, Bitbucket, Azure DevOps	Qualification required. Free PR summaries are available more broadly
Qodo	”Qodo for Open Source” programme	Apply	GitHub, GitLab, Bitbucket, Azure DevOps	No permanent free tier of any kind otherwise
Ito	Unlimited public repos, runtime analysis on every PR, video and screenshots per run, for qualified non-commercial MIT or Apache projects	Apply	GitHub only	Licence list is narrower than “OSI-approved”. No GitLab
CodeAnt AI	”100% off for open source” on request	Apply (email)	GitHub, GitLab, Bitbucket, Azure DevOps	Not a self-serve path
GitHub Copilot	Copilot Pro free for verified students, teachers and maintainers of popular open source repos, which includes Copilot code review	Verification required	GitHub	Copilot Free does not include automated PR review, only “Review selection” in VS Code

Two findings in that table are worth pulling out.

Korbit giving away its top tier to open source is the most generous single offer we found, and almost nobody writes about it. Korbit Max is free for all open source repositories with no published cap, while paid Max is $18 per user a month billed annually. The absence of a published cap is both the appeal and the risk.

The apply-first group is where evaluation timelines die. Greptile, Gitar, Qodo, Ito and CodeAnt AI all require you to qualify before you get free open source access. If you are a maintainer with a weekend to spend, start with the install-it-yourself group: CodeRabbit, Sourcery, DeepSource, Korbit, cubic and Codacy all go live in minutes.

AI code review tools with a free plan

Direct answer: 9 of the 51 tools we track have a permanent free plan that works on private code, and all 9 are capped. cubic Starter: 20 PR reviews/mo per GitHub org. Greptile Starter: 1 active dev, 50 credits/mo. CodeRabbit Free: 1 private PR review/hour, summary only. Codacy Developer: IDE-only. Semgrep: 10 contributors, 10 private repos. Verified 2026-10-18.

This is category 2 from the framing above: free with a hard usage cap, but it works on private code. It is the group that matters to most readers and the one most roundups get wrong, because they list public-repo offers in the same column.

Read the cap column twice. It is where the free plan ends.

Tool	Free plan name	Works on private code?	The hard cap	What is excluded	Card required
cubic	Starter	Yes	20 PR reviews/mo per GitHub organisation	The CLI, so no local or CI-pipeline reviews	No
Greptile	Starter	Yes	1 active developer, 50 credits/mo, unlimited repos	Base review = 1 credit, Plus = 3, Apex = 10, so depth burns the allowance fast	No
CodeRabbit	Free	Partially	1 PR review/hour, summary only, plus 3 IDE and 3 CLI reviews/hour	Full inline review on private code	No
Codacy	Developer	Yes	IDE plugin only, no cap on usage	Cloud-hosted scans, PR decoration, AI Reviewer, merge gates	No
Semgrep	Free Edition	Yes	10 contributors, 10 private repos, 60 AI credits	Higher repo caps, Secrets product	No
SonarQube Cloud	Free	Yes	5 users, 50,000 private lines of code; unlimited public projects	AI CodeFix is paid-tier. Analysis that would exceed the LOC limit is simply not performed	No
Snyk Code	Free	Yes	5 projects, 100 Snyk Code tests/mo	Jira integration, higher limits	No
Qlty	Cloud Free	Yes	Unlimited public and private repos and contributors, but 1,000 analysis minutes and 100 AI autofixes/mo	Monorepo support, REST API, accelerated analysis	No
Graphite	Hobby	Personal-account repos only	”Limited AI Reviews”, number not published	Organisation repos, Graphite Chat limits	No
Amazon Q Developer	Free Tier	Yes	50 agentic requests/mo, 1,000 LOC/mo Java transforms	Admin dashboard, IP indemnity	No
Qodana	Community	Yes, self-hosted	Unlimited lines of code, limited linters: JVM, Java for Android, Python, .NET, C, C++	Framework support, most paid features, long data retention in Qodana Cloud	No
GitHub Copilot	Free	Yes, but not for PR review	~2,000 completions and 50 chat requests/mo, auto model selection only	Automated pull request review. Only “Review selection” in VS Code	No

The three honest conclusions from that table

1. cubic Starter is the most usable free plan for a small team on private code. 20 pull request reviews a month per GitHub organisation, no card, with custom agents, automatic PR descriptions and unlimited AI wikis included. The CLI exclusion is the real limit: you cannot run cubic locally or in CI on the free plan. Note that cubic’s own site and blog say 20 and several third-party directories say 40; we report 20 because it is vendor-published.

2. Greptile Starter is a solo-developer plan, not a team plan. 1 active developer and 50 credits a month sounds generous until you read the credit table: a base review costs 1 credit, a Plus review 3, and an Apex review 10. Run Apex depth and the monthly allowance is five reviews. Greptile also defines a seat as any developer who received a review in the billing period, and credits are per author and do not pool, so the model does not get more forgiving as you grow. Detail in Greptile pricing.

3. CodeRabbit’s free tier changed during 2026 and most of the internet has not caught up. The current pricing page states free reviews forever for public repositories and does not include private repositories on the free plan. What free accounts keep on private code is 1 summary-only PR review an hour, 3 IDE reviews and 3 CLI reviews an hour. Plenty of 2026 articles, including an older post on this site, still describe unlimited free private repo reviews. That is stale. The genuinely useful survivor is the CLI, which reviews your local working tree free with an account at 3 reviews an hour. See is CodeRabbit free? and CodeRabbit pricing.

The free plan that is not a reviewer at all

Two entries above deserve a warning label. Codacy’s Developer tier is free forever and genuinely good, but it is an IDE product. 38 languages, SAST, SCA and secrets scanning, scan-as-you-type, auto-fix handoff to an agent, in VS Code, JetBrains and Cursor. No cloud scans, no PR decoration, no AI Reviewer, no merge gates. If you want findings while you type, it is the best $0 in the market. If you want a bot on your pull requests, it is not that. Codacy free vs Pro has the full split.

GitHub Copilot Free does not include automated pull request review. It gives you completions and chat plus “Review selection” inside VS Code. Copilot code review is a paid-plan feature, and on paid plans it carries a cost most buyers miss: since 1 June 2026, code review on private repositories consumes GitHub Actions minutes in addition to AI credits, and GitHub publishes a model multiplier of 13 for code review against the premium requests meter. Private repos get 2,000 included Actions minutes a month on Free and 3,000 on Pro and Team. See GitHub Copilot pricing.

Category 3: genuinely free because open source and self-hosted

This is the only category with no cap. It is also the only category where “free” is a slightly dishonest word, and we will do the arithmetic below.

Tool	Licence	Verified state	Deployment	Models	The honest caveat
PR-Agent	MIT (Aug 2026)	Transferred from Qodo to a community organisation; ~13.1k GitHub stars; releases roughly biweekly, latest on PyPI 2 Oct 2026	CLI, Docker, GitHub Action, or webhook server	OpenAI, Claude, Gemini, or local via Ollama and LiteLLM	README describes it as a community-maintained legacy Qodo project and states plainly that it is not the Qodo free tier. Two sources list it as Apache-2.0 rather than MIT; PyPI metadata says MIT
Kodus	AGPLv3 (Sept 2026), except the ee/ directory which is commercially licensed	selfhosted-2.2.4 released 27 Sept 2026	Docker Compose on one VM, Kubernetes or OpenShift via Helm. No seat minimums, no sales call	12 built-in providers plus any OpenAI-compatible API, including vLLM, Ollama, TGI and LiteLLM	SSO, RBAC and audit logs sit behind a commercial licence key, not in the AGPL build. Self-hosted instances send one anonymous daily heartbeat of aggregated counters, disableable via an environment variable
SonarQube Community Build	Open source, self-hosted	5,000+ rules, 30+ languages	Self-hosted	Deterministic analysis, no LLM	Analyses the main branch only. Pull request analysis is not included. It is a quality gate, not a PR reviewer. AI features are Enterprise-tier
Qlty CLI	Fair source, BSL 1.1 with delayed open publication	Rust, no Docker dependency, Mac/Linux/Windows	Local or CI	70+ static analysis tools across 40+ languages	Free for all use including commercial with no contributor limits, but fair source is not open source today. Call it what it is
Tabby	Apache 2.0, except the ee/ directory	v0.32.0 released 25 Jan 2026, ~33k stars	Fully self-hosted, air-gap capable	Local models on your own GPU	Tabby is a completion server first. PR review and diff analysis are outside its core scope. Community edition covers up to 5 users; SSO is in the enterprise directory. Needs GPU compute capability 7.0+; CPU inference at 8-15 tokens/second is not usable
Vercel OpenReview	Open source, self-hosted	Beta	Deploy to Vercel, connect a GitHub App, trigger by mentioning the bot in a PR comment	Your own keys	Explicitly a Vercel internal project released in beta, with expected rough edges and breaking changes

What “free” actually costs when you self-host

Here is the part the open source roundups skip. You do not eliminate the cost, you move it. A self-hosted reviewer replaces a per-seat licence with three line items:

1. Infrastructure. A webhook server and a worker if you bring your own API key. A GPU host if you run a local model.
2. Model inference. Either your provider’s list price per token, or GPU time.
3. Engineer time. Setup, upgrades, incident response, and the ongoing cost of being the person who owns it.

One 2026 total-cost analysis puts GPU spend at 60 to 70 percent of self-hosted TCO and engineer time at 25 to 30 percent. Read that again: in a self-hosted configuration, the licence fee you avoided is usually the smallest number in the comparison. The cheapest honest configuration for most teams is therefore self-host the reviewer, bring your own API key for the model - PR-Agent and Kodus both support this, and it collapses your infrastructure cost to a small server while inference goes to your provider at list price with no vendor markup.

On local-model review quality, we are going to be careful

You will see a claim circulating that an air-gapped local-model setup delivers roughly 60-70% of cloud-tier review quality. We could not verify that as a published measurement of review quality, and we are not going to repeat it as one.

What we actually found, on 2026-10-18:

• An independent blog benchmark of 50 coding tasks on a $489 consumer GPU concluded that a local Qwen3-Coder setup can replace a frontier model for roughly 60-70% of daily coding work - single-file generation, unit tests, quick bug fixes, explanations, documentation - scoring within 6% on average, with the gap concentrated in multi-file refactoring and architectural reasoning, plus 30-40% more time spent reviewing the output. That is a judgement about task mix, not a quality ratio, and it is about coding rather than reviewing.
• A separate 2026 TCO analysis recommends a router keeping 60-70% of traffic local, and explicitly assigns first-pass code review to the local tier, quantifying the capability gap as 4-8 months behind frontier hosted APIs.
• The clearest sourced statement about review specifically is qualitative: open models have closed much of the gap, but review depth on subtle, cross-file logic issues still correlates with model strength.

So: the “60-70%” number is real, it appears twice in this literature, and in neither case does it mean “60-70% of cloud review quality”. Two different things got collapsed into one statistic. Treat it as a planning heuristic for task routing, not as a benchmark result.

One independent finding that is about review and is worth your attention: a 2026 arXiv study of self-review failures in LLM-based code modernisation found that across 262 semantic drift cases, 83 (31.7%) were silently endorsed as behaviour-preserving by the same model that produced them. The self-miss rate was bimodal - five models missed essentially none of their own drift, five missed between 20% and 100% - and it was not predicted by the drift rate. The practical conclusion is that self-review reliability is a per-model property, so validate the specific model you plan to run rather than assuming any general local-versus-cloud percentage transfers.

For the wider independent evidence base, the c-CRAB benchmark from the National University of Singapore (arXiv 2603.23448) converted human review comments into executable tests and found review agents collectively identify only about 41.5% of issues - a finding about the whole category, cloud models included. Note that the frequently quoted “40-70% false positive rate” attributed to c-CRAB does not appear in that paper. We cover the evidence landscape in the state of AI code review in 2026 and the self-hosting trade-offs in open source AI code review tools. One disclosure, because we hold other publishers to it: c-CRAB is not conflict-free. Five of its six authors are academic (National University of Singapore, Zhejiang University), but one lists SonarSource as their affiliation (arXiv 2603.23448v3, verified 2026-10-18), and the paper carries a disclaimer that its conclusions do not represent SonarSource policy. Sonar sells SonarQube and acquired Gitar, which is a paid sponsor of this site. SonarQube is not among the four tools c-CRAB evaluates, which blunts the conflict considerably - but we are not going to call out vendor-funded benchmarks elsewhere and stay quiet about this one.

Best AI code review tools for open source projects

Direct answer: CodeRabbit (free forever on public repos, four forges), Sourcery (public GitHub and GitLab, with auto-approve for low-risk PRs), DeepSource (1,000 public PRs/mo plus a 5,000-rule deterministic pre-pass) and Korbit Max (free on all open source repos) are the four strongest install-it-yourself options. Greptile, Gitar, Qodo, Ito and CodeAnt AI are free for OSS but require an application. Verified 2026-10-18.

Criteria: depth of review, breadth of forge support, and whether you can turn it on today without emailing anyone.

CodeRabbit - the most capable, and the loudest. Free reviews forever on public repositories, subject to a fair-usage policy, and the only option in this group covering GitHub, GitLab, Bitbucket and Azure DevOps. It reads .coderabbit.yaml, so you can scope paths and suppress categories. Do that before you enable it on a busy repo.

Sourcery - the best choice if your constraint is your own time. The free public-repo plan includes AI reviews, PR summaries, diagrams, one-click fixes, custom review rules and auto-approval of low-risk PRs. That last feature is the one that actually reduces maintainer workload rather than adding to it. GitHub and GitLab only. Sourcery pricing covers the paid tiers.

DeepSource - the best signal-to-noise for a contributor-facing repo. Unlimited public repos and 1,000 pull requests reviewed a month. The architectural reason to care: DeepSource runs deterministic static analysis across 5,000+ rules and 30+ languages before the AI agent sees the code, so lint and style findings arrive as facts rather than model guesses. That produces fewer arguments in comment threads with contributors you have never met. The catch is that AI Review and Autofix are pay-as-you-go on top at $8 per 10,000 lines standard and $15 advanced, so the fully free configuration is the deterministic layer. See DeepSource free vs paid.

Korbit - the most generous tier, with the least documentation. Korbit Max, its top non-enterprise tier at $18 per user a month paid, is free for all open source repositories with no published cap. GitHub, GitLab and Bitbucket. No published cap means no published guarantee, so do not build a release process on it.

The apply-first alternatives, honestly assessed. Greptile and Gitar both require an OSI-approved licence. Ito requires MIT or Apache specifically and is GitHub only, but offers something genuinely different: it builds and runs your app in an ephemeral environment and returns runtime evidence with video and screenshots per run, at the cost of a 45-to-60-minute review. Qodo has no free tier at all outside its open source programme. CodeAnt AI handles open source by email.

Disclosure, stated here rather than in a footer: Gitar is a paying sponsor of aicodereview.cc and Ito is an unpaid editorial baseline partner. Neither is ranked above the install-it-yourself group, because requiring an application is a real disadvantage for a maintainer and no sponsorship changes that. We will also note, as we do everywhere review quality comes up, that neither Gitar nor Ito has a published third-party benchmark. See our Gitar review.

Best AI code review tools for open source maintainers

Direct answer: Maintainers should optimise for contributor experience, not review depth. Sourcery’s auto-approve for low-risk PRs cuts load directly. DeepSource’s deterministic pre-pass reduces arguments with drive-by contributors. CodeRabbit is the most capable but the most verbose, so configure before enabling. Maintainers of popular repos may qualify for free GitHub Copilot Pro, which includes code review. Verified 2026-10-18.

A project and a maintainer are not the same buyer, which is why we are answering this twice. A project wants the best review. A maintainer wants fewer hours spent on pull requests from strangers, and a bot that does not drive those strangers away.

The maintainer-specific checklist, in order of how much pain each item causes:

1. Can you scope it to paths? A reviewer commenting on docs/, vendor/ and generated files is a net negative. CodeRabbit’s .coderabbit.yaml and PR-Agent’s .pr_agent.toml both support this. Check before enabling.
2. Can you separate summaries from inline findings? Some maintainers want a summary and no inline noise; some want the inverse. A tool that only does both is a tool you will turn off.
3. Does the free tier survive a traffic spike? “Fair usage policy” is a phrase that resolves itself during Hacktoberfest. CodeRabbit’s public-repo free tier is explicitly subject to one. Nobody publishes the threshold.
4. Does it support your forge? CodeRabbit covers GitHub, GitLab, Bitbucket and Azure DevOps. Sourcery covers GitHub and GitLab. cubic, Ito, Macroscope and Vercel OpenReview are GitHub only. Self-hosted Kodus adds Forgejo, which is the only option here for a project that has left GitHub on principle.
5. Does your licence actually qualify? Greptile and Gitar require OSI-approved licences. Ito specifies MIT or Apache. A source-available licence can disqualify you from a free tier you assumed you had. Read the qualification wording before you plan around it.

The two routes most maintainers miss. First, if you maintain a popular open source repository you may qualify for GitHub Copilot Pro free, which includes Copilot code review - unlike Copilot Free, which does not. Verified students and teachers qualify on the same basis. Second, self-hosting is more viable for a maintainer than for a company, because you already have CI and you do not need SSO, RBAC or audit logs. PR-Agent under MIT as a GitHub Action with your own API key costs you a secret and a workflow file, and nothing else is gated.

What we would actually do, maintaining a mid-size public repo today: Sourcery for auto-approval of low-risk PRs, DeepSource’s deterministic layer for style and lint so those findings never become a debate, and PR-Agent on demand via a slash command for the PRs that need real depth. That combination is $0 and configurable, and none of it requires an application.

What you do not get for $0

Stated plainly, because every free tier in this market withholds roughly the same things:

• Identity and governance. SSO/SAML, custom RBAC and audit logging are Enterprise entitlements at CodeRabbit, Qodo and Codacy. In Kodus they sit in the commercially licensed ee/ directory, and in Tabby in the enterprise directory - so even the open source options gate them.
• Self-hosting of commercial tools. CodeRabbit requires Enterprise and 500 or more seats. Qodo puts on-prem and single-tenant on Enterprise. DeepSource puts self-hosted on Enterprise. If self-hosting is a hard requirement, your real shortlist is category 3.
• Support with an SLA. Free tiers get community support. Qlty’s free plan is explicit that it is Discord and GitHub on a best-effort basis.
• Rate headroom. The free plan is where the rate limits live: 1 private review an hour at CodeRabbit, 20 a month at cubic, 50 credits at Greptile, 1,000 analysis minutes at Qlty.
• AI features on static-analysis tools. SonarQube Cloud’s free tier excludes AI CodeFix. SonarQube Community Build excludes AI features entirely and does not do PR analysis at all. See is SonarQube free?
• Monorepo support. Qlty Cloud’s free plan excludes it, which matters more than it sounds if your repo layout is non-trivial.
• The CLI. cubic’s Starter plan excludes it, so no local or CI reviews.

Free tier or free trial - which do you actually want?

These are different purchases, and most roundups blur them into one list.

Use a free tier when you have no budget, or when you need longer than two weeks. A free tier does not expire mid-sprint. cubic Starter, Greptile Starter, Codacy Developer, Semgrep Free Edition and SonarQube Cloud Free all let you move at your own pace.

Use a free trial when budget is approved and you are choosing between two finalists. 14 days is the market standard and seven vendors require no credit card. The data - trial length, card requirement, review quota, what is gated and exactly what happens on day 15 - is in AI code review tools with a free trial.

Use a credit grant when you want neither a clock nor a cap. Macroscope gives every new workspace $100 of usage at $0.05 per KB of diff reviewed, DeepSource bundles up to $50 of AI Review credits, and Ito makes your first 20 reviews free. These are one-time grants, not permanent tiers, which is why they are not in the tables above.

How we verified this

What we checked. On 2026-10-18 we read the live pricing pages and, where published, the product documentation for CodeRabbit, Sourcery, DeepSource, Korbit, cubic, Codacy, Macroscope, Panto, Greptile, Gitar, Qodo, Ito, CodeAnt AI, GitHub Copilot, Semgrep, SonarQube Cloud and Community Build, Snyk, Qlty, Graphite, Amazon Q Developer, Qodana, PR-Agent, Kodus and Tabby. For each we recorded whether the free offer is permanent, whether it touches private code, the exact cap, and what is excluded.

Primary sources. CodeRabbit pricing and rate limit docs, Sourcery pricing, DeepSource pricing, Greptile pricing, Codacy pricing, Semgrep pricing, Snyk plans, Graphite pricing, Qodo pricing, Qodana editions, the PR-Agent PyPI listing for licence metadata, and Kodus self-hosted docs.

Conflicts we found and how we resolved them.

• CodeRabbit’s free tier. The pricing page states free reviews forever for public repositories and shows private repositories as not included on the free plan. Numerous 2026 third-party articles, and one older post on this site, describe a free tier with unlimited private repo reviews. We report the current pricing page and docs, and we are flagging our own stale post rather than quietly leaving it.
• cubic’s free review count. cubic’s own site and blog say 20 PR reviews a month per GitHub organisation. Some third-party pricing snapshots say 40. We report 20 as the vendor-published figure and note the conflict.
• cubic’s paid prices. The live pricing page shows Team $40, Pro $99 and Max $200 per developer a month with a 20% annual discount. Other sources cite Team at $30 annual. We report the live figures.
• PR-Agent’s licence. PyPI metadata says MIT, with the latest release dated 2 October 2026, and the community transfer is documented. Two secondary sources describe it as Apache-2.0. We report MIT and flag the conflict; check the LICENSE file in the community repo if the distinction matters to your legal review.
• The “60-70%” local-model figure. Addressed at length above. The number exists in two different 2026 analyses, in neither case as a measurement of review quality. We decline to report it as a review-quality benchmark.
• Anthropic’s Code Review credit grant. Our own internal research note recorded $10,000-plus in review credits for eligible Enterprise organisations; live coverage of the GA announcement reports up to $50,000. We have not included Claude Code Review in this article’s tables because it has no free tier at all - it is unavailable on Claude Free and Pro - but the discrepancy is worth recording. See Claude code review.

What we could not verify.

• Graphite’s Hobby AI review limit. The pricing page says “Limited AI Reviews” with no number. A Graphite guide elsewhere on their site states the Agent is free for up to 100 pull request reviews a month, but we could not confirm that figure applies to the Hobby plan. Unverified.
• Whether Semgrep’s 60 AI credits are monthly or one-time. The pricing page lists “60 AI credits” for Free Edition against “20 per developer per month” for Teams, which reads as a one-time grant. We could not confirm. Unverified.
• Korbit Max’s open source cap. Korbit publishes no cap for open source repositories. Absence of a published cap is not a guarantee of no cap. Unverified.
• cubic’s public-repo review limit. cubic states it is free for public repositories with no limit mentioned. Whether an unstated fair-usage threshold applies is unverified.
• CodeAnt AI’s full plan table. The pricing page confirms 100% off for open source via email, but the plan comparison table did not render in our fetch. Unverified on 2026-10-18.
• Panto’s review-product pricing. Panto’s open source offer is confirmed on its GitHub App listing, but per-seat pricing for the code review product is not publicly posted. Third-party trackers disagree across a $12 to $50 range. Unverified.
• Gemini Code Assist’s free individual tier. One source reports the individuals tier stopped serving requests on 18 June 2026 with users redirected to Antigravity; another 2026-dated page still advertises a free tier. Google’s own quota docs state that a GitHub app installation has a quota of at least 100 pull request reviews a day, tracked separately from IDE quotas. We could not resolve whether the free individual tier still exists and have excluded Gemini Code Assist from the tables rather than guess. If it does still exist, it would be the most generous free private-repo offer in this market by an order of magnitude, so it is worth checking yourself.
• Review quality on any free tier. No independent benchmark evaluates free tiers specifically. The nearest independent evidence, c-CRAB and the Entelligence Code Review Benchmark, tests the products, not the plans.

Our conflict of interest. Gitar is a paying sponsor of aicodereview.cc and Ito is an unpaid editorial baseline partner. Both appear in the apply-first group above, ranked below the install-it-yourself options, because that is where the evidence puts them for this use case.

Recommendations

Criteria: works on private code without an application, has a cap you can actually live with, and does not require you to run infrastructure unless you chose to.

Best free plan for a small team on private code: cubic Starter. 20 pull request reviews a month per GitHub organisation, no card, no application, with custom agents and automatic PR descriptions included. For a team shipping 5 PRs a week this is a genuinely usable permanent free tier, which almost nothing else in this market is. Two limits to go in knowing: no CLI on the free plan, and GitHub only.

Best free plan for a solo developer: Greptile Starter. 1 active developer, 50 credits a month, unlimited repositories. Keep reviews at base depth - 1 credit each - because Plus costs 3 and Apex costs 10, and at Apex depth your month is five reviews.

Best $0 for findings while you type: Codacy Developer. Free forever, 38 languages, SAST, SCA and secrets, in VS Code, JetBrains and Cursor. Understand what you are getting: this is an IDE tool, not a PR reviewer.

Best free tier for a public repository: Sourcery if you are short on time, CodeRabbit if you want depth, DeepSource if your repo gets drive-by PRs. Sourcery’s auto-approve for low-risk PRs is the single most maintainer-friendly feature in the free category. CodeRabbit is the most capable and covers four forges, but configure .coderabbit.yaml before you enable it. DeepSource’s deterministic pre-pass across 5,000+ rules means style findings are facts, not model opinions, which is worth a lot in a contributor thread.

Best genuinely free with no cap: PR-Agent, or Kodus if you want a UI. PR-Agent is MIT as of August 2026, runs as a CLI, Docker image, GitHub Action or webhook server across five forges, and works with OpenAI, Claude, Gemini or local models. Kodus is AGPLv3 as of September 2026, self-hosts free at any team size via Docker Compose or Kubernetes, adds a web dashboard, review history and custom rules, and supports Forgejo. Run either with your own API key rather than a local GPU unless you have an air-gap requirement - that configuration gives you the best quality-per-dollar, because GPU spend is 60 to 70 percent of self-hosted total cost and you are unlikely to beat a frontier model on review depth with hardware you can afford.

What we would not do. We would not pick a free tier on the assumption that review quality is the same as the paid tier - it usually is, but the rate limits are not, and a reviewer you have to wait an hour for is a reviewer your team routes around. And we would not plan a budget around an unpublished cap. Korbit’s uncapped open source tier and cubic’s unlimited public repos are both great offers and neither is a commitment.

If you need commercial terms rather than $0, the trial data is in AI code review tools with a free trial. For the shortlist ranked on review-quality evidence rather than price, see the best AI code review tools, best AI PR review tools and what is AI code review.