Prelint

Prelint is a product-intent review layer for pull requests. Its core distinction is explicit: conventional code review asks whether an implementation works; Prelint asks whether the change matches what the team previously decided to build.¹² It reads product specifications, architecture decision records, business rules and related repository context, then flags drift, contradictions, duplicated decisions, compliance conflicts and business-logic changes directly on the PR.²³

That focus matters because the failure mode is different from syntax or vulnerability detection. A change can compile, pass tests and look clean while still changing pricing logic, introducing a vendor that conflicts with an architecture decision, weakening a retention rule or building an off-roadmap API. Prelint is designed to catch that class of mismatch rather than compete with linters, type checkers or security scanners.¹ Its own review documentation explicitly says it does not target style, formatting, type-safety or security-vulnerability scanning and should sit beside tools dedicated to those jobs.³

The GitHub review pipeline is concrete. A PR open/reopen/ready event or a new commit triggers a webhook; Prelint validates the event, assembles product context, runs an AI review against the diff, then runs a second AI validation pass that filters findings which look like linter territory, speculation, contradictions with documented context or claims without diff evidence.³ Findings are posted inline with a summary, and the system can automatically approve a PR when no active product-intent findings remain.³

Context management is central to the product. Current docs describe an approximately 80K-character review-context budget, prioritized from more specific to more general material; when a repository clone is available, the review agent can also search the broader codebase.²³ This is a sensible defense against indiscriminately stuffing an entire organization history into every review, but it creates an important dependency: Prelint can only enforce decisions that are documented clearly enough to retrieve and interpret. Stale or contradictory ADRs can therefore become an input-quality problem rather than a model-quality problem.

The product increasingly extends beyond the PR bot. Current official surfaces describe GitHub, CLI and MCP workflows so coding agents can query product decisions while they work rather than discovering drift only at final review.¹⁴ That direction is strategically strong: if an agent can ask what the approved pricing model or infrastructure standard is before editing code, Prelint can become preventive context infrastructure rather than merely a post-hoc reviewer. The independent exact-product evidence for these newer interfaces is still thin, however.

Prelint also treats repeated review as a real event rather than an unlimited hidden operation. New commits or force pushes can trigger a fresh review, with current billing charging each completed re-review.⁵⁶ CI bots such as Dependabot-style automated accounts are excluded from paid triggering, while ordinary AI coding agents are treated like normal authors.⁵ This is economically legible but makes agent behavior relevant to cost: noisy iterative PRs can accumulate review charges even when a human sees them as one logical task.

The project publishes a first-party research program called AI Code Pulse. Its current research page says it analyzed 56,706 pull requests across 331 open-source repositories, grading PRs with and without documentation context and reporting 80.8% precision on context-driven flags, 979 catches visible only with context and 232 false alarms suppressed by context.⁷ It also reports that review accuracy improved substantially with documentation and that overzealous flags were a larger failure mode than hallucinations.⁷ This is more useful than an unsupported homepage percentage because the methodology and component results are partially exposed, but it remains vendor-run research rather than an independent benchmark of production Prelint deployments.

Independent research supports the broader problem without validating Prelint specifically. METR found that many AI-generated patches which pass SWE-bench-style automated grading would not be accepted by maintainers and argues that algorithmic benchmark success maps imperfectly to real-world mergeability.⁸ That strengthens the case for richer review layers, but METR evaluated maintainability/quality dimensions and did not test Prelint’s product-intent accuracy. It should therefore be read as problem-space evidence, not a product score.

Prelint’s current commercial model is unusually simple: $1 per completed review.⁵⁶ There are no seat fees, subscription minimums or repository charges in the current pricing documentation. Signup currently includes $10 of free credits without requiring a card, giving roughly ten completed reviews before a paid top-up is needed.⁵

Credits are prepaid and described as non-expiring.⁵⁶ Failed, canceled or timed-out reviews are not charged, and public/open-source repository reviews are currently free.⁵ Re-reviews after material PR updates are charged again at $1 per completed run. For a team with disciplined PR boundaries, this is easier to forecast than per-seat pricing; for an agentic workflow that pushes many iterations, review count becomes the relevant cost driver.

Auto-recharge is enabled by default on the current billing surface, with a default threshold of $10 and a $20 top-up; users can change or disable it.⁵⁶ Monthly spending limits can also pause reviews when a configured cap is reached. This is a good cost-control mechanism, but teams should configure it deliberately rather than assume prepaid credits alone prevent automatic spending.

The current top-up interface supports purchases from $10 through $1,000, and larger organizations can ask about invoices or annual prepayment.⁵ Prelint also advertises startup credits and referral credits; these are promotional benefits rather than core unit economics and should not be used to compare long-term cost.

The major pricing trust issue is legal-document drift. Prelint’s live pricing and billing docs repeatedly say there is no subscription and charges are prepaid per completed review.⁵⁶ The Terms, effective July 30, 2026, still state that paid plans are billed in advance monthly or annually, discuss renewal-style pricing changes and describe free trials as discretionary.⁹ Those provisions materially describe a different commercial model. The product may operationally bill exactly as the current app/docs say, but the contract should be updated so a buyer does not have to choose between the operative UI and the governing public Terms.

Capterra illustrates how quickly the commercial model changed. Its exact-product page was updated August 26, 2026 but still lists Prelint at $30 per unique committer per month with unlimited reviews.¹⁰ That conflicts with current first-party $1-per-review pricing, so Capterra’s commercial fields are stale and should not be blended with the live billing model. The Capterra page has zero user reviews, so it provides no user-rating signal.

Value depends strongly on what Prelint catches. At $1, even a modest number of genuinely prevented product mistakes would make the tool inexpensive relative to engineering time. Conversely, if a repository has weak documentation or produces mostly low-value/false-positive comments, low unit price does not rescue poor signal. Without an independent production benchmark of accepted/fixed findings per paid review, value-for-money remains unscored rather than inferred from the headline dollar amount.

Independent exact-product evidence remains sparse for Prelint. Capterra has zero reviews and stale commercial data relative to the live product.¹⁰ No mature G2/TrustRadius-style user corpus or independent exact-product benchmark was identified. Missing evidence is not negative evidence, but it limits confidence in comparative quality and reliability claims.

Product Hunt currently shows 5.0/5 from three reviews, about 1.3K followers and #1 Product of the Day and Week for the July 29, 2026 launch.⁴ The most useful review describes Prelint as effective at catching product decisions made by coding agents and aligning them with ADRs, while asking for stronger first-run/product-team interfaces; it also says the reviewer uses Prelint alongside Greptile because the tools cover different classes of problems.⁴ That is directly relevant hands-on signal, but three reviews are too few and too launch-adjacent to support a stable numeric product-quality score.

Product Hunt also carries several maker claims, including that roughly 40% of fixed-before-merge comments on teams using multiple AI reviewers come from Prelint and that usage grew roughly 10x month over month after launch.⁴ These are useful hypotheses for future verification but remain first-party claims. They are not converted into independent adoption or accuracy inputs.

Security documentation is a relative strength. Prelint says it never trains models on customer code, uses per-organization isolation, encryption at rest and in transit, least-privilege access and audit logging.¹¹¹² The security page says AI inference runs through AWS Bedrock under zero-retention terms and that source code does not leave AWS in that architecture.¹¹ It also says dedicated review containers are destroyed after work completes, repository storage is isolated and access by employees is restricted/logged.¹¹²

Those are substantial first-party controls, not independent certification. Prelint explicitly says SOC 2 Type I is in progress, with Type II observation to follow; it must not be described as SOC 2 certified today.¹¹ The security page also mentions external infrastructure review and enterprise options such as dedicated infrastructure/data-residency arrangements, but these do not substitute for a completed Prelint assurance report.

GitHub permissions are documented in useful detail. Current security material explains read access to repository metadata and content, pull-request and checks permissions, organization-membership read access, and optional write permission for suggested-fix commits; inbound GitHub webhooks are validated with HMAC-SHA256.¹¹ This level of specificity is valuable for engineering-security review because it exposes the actual blast radius rather than merely saying “least privilege.”

Privacy documentation is reasonably clear about categories but still leaves retention questions. The Privacy Policy says code diffs and metadata are temporarily processed, account data is retained while the account is active, deletion follows account deletion within up to 30 days, and review metadata may be retained for audit logging and service improvement.¹² It also says relevant diffs, PR metadata and product specifications may be sent to AI providers under contractual zero-retention terms.¹² The security/docs pages provide stronger implementation statements, but the public privacy text does not give a simple category-by-category duration for active repository clones or review records.

The product sensibly frames AI review as advisory. Its Terms state that AI-generated output is not guaranteed to be accurate and human judgment remains necessary.⁹ This limitation aligns with Prelint’s documented two-pass validation approach and with its own research finding that overzealous review can be a significant failure class.³⁷ Teams should measure accepted/fixed finding rates on their own repositories rather than assume that a sophisticated review pipeline eliminates false positives.

The broader independent research context reinforces that caution. METR’s maintainer-review work found substantial gaps between automated test success and real-world merge acceptance for AI-generated patches.⁸ It supports the need for layered human/automated governance, but it also means Prelint should be evaluated as one control in a stack: product-intent review, conventional code review, CI/tests and security scanning address different failure modes.

Prelint is best suited to engineering organizations where product intent is actually written down. Teams maintaining ADRs, product specs, compliance rules, pricing logic and architecture constraints give the reviewer a meaningful source of truth.² If important decisions live only in meetings or individual memory, the system cannot reliably enforce them; adopting Prelint may therefore expose documentation debt before it delivers full review value.

Teams using coding agents heavily are the clearest audience. Agentic coding increases change volume and can produce plausible implementations faster than humans can re-check every historical decision. Prelint’s ability to review each PR and expose the same decision context through CLI/MCP can add a product-governance layer without forcing a human product manager into every iteration.¹⁴

Regulated or policy-heavy software can also benefit when compliance constraints are expressed as reviewable repository context. Prelint explicitly positions product rules and compliance decisions as review criteria.¹ That does not make it a compliance certification tool: the system can flag divergence from documented policy, but legal correctness still depends on the quality of that policy and human review.

Small teams may find the per-review model attractive because there is no seat tax.⁵ A five-person team and a fifty-person team pay according to completed review volume rather than headcount. Public/open-source projects have an especially low-friction path because public-repository reviews are currently free.⁵

Prelint is less suitable as a replacement for conventional code-review platforms. Its own docs deliberately exclude style, type safety and vulnerability scanning.³ Teams expecting one bot to catch product drift, implementation bugs, dependency vulnerabilities, test gaps and formatting issues will need additional tooling or humans.

It is also a weaker fit for organizations that require completed independent assurance before connecting source repositories. The current security design is detailed and directionally strong, but SOC 2 Type I is still in progress.¹¹ Enterprises with formal procurement requirements should request the latest audit status, DPA/subprocessor details and retention architecture before production rollout.

Finally, high-volume autonomous-agent teams should monitor both signal quality and spend. At $1 per completed review, raw price is low, but many small PR revisions can create many billable re-reviews.⁵ A useful deployment metric is not just “reviews run”; it is cost per accepted product finding, false-positive rate, time saved and number/severity of prevented drift events.