Open the riskiest PR first. Know what to check before you read a line. Read what a PR does in plain English. Ask the PR questions, get answers from the diff. See what a change could break before you merge. Find out if your tests would catch it. Fast-track low-risk PRs with approvals you control. See where review is slowing your team down. Merge with confidence, not crossed fingers. Keep your source code and your keys yours.
Free plan · 1 watched repo · Bring your own AI key
The problem
More pull requests land every sprint than you can reasonably get through, and AI-assisted coding tools are only pushing diff volume higher: bigger changes, more often, with the same number of people looking at them. Every PR still gets the same manual once-over regardless of how risky it is, so low-risk changes eat the attention the risky ones need.
Before: every PR looks the same, so you open them in arrival order.
Risk score playground
The same weighted formula Qodillo runs on every PR — no LLM involved in the number.
Points are each signal's live share of the 0–10 total. Review depth is held at a resolved-discussion baseline (0.28 pts, ×1.0).
AI missions · preview
Real missions are generated per PR (3–6, sorted High → Medium → Low) using your stack's review standards.
Features
Prioritize
A deterministic 0–10 risk score for every PR — CI status (×2.5), security files (×2.0), file volume and line changes (×1.5 each), review depth (×1.0), merge timing (×0.5). A formula, not an LLM guess.
A live queue of the PRs you authored or are asked to review across your watched repos, each with its risk score.
Approval and merge trends over time, so you can see where review is actually slowing your team down.
Paid plans: see how much review pressure each repository has, with a daily history. It uses no AI and costs no usage.
Pro and Team, off by default: score every opened or updated PR against your risk policy, with an optional AI judge double-checking only the lowest-risk tier. Posting an approve, comment or request-changes review to GitHub is a separate opt-in.
Review
AI-generated review missions, prioritized High to Low, tell you exactly what to look at and why — shaped by stack-specific standards for Java, Spring Boot, React, Python, TypeScript and more.
A multi-turn AI chat scoped to the PR — ask follow-up questions without leaving the page and get answers grounded in the real diff, CI and comments.
A bounded tool-calling loop — three rounds of real tool calls, then one forced synthesis round — checks test coverage, CI status, changed files, and coding patterns automatically before you ask.
Pro and Team: write organization-wide review guidance in plain language, or ship your own Python skill that runs in an isolated sandbox against every PR diff. Every skill is safety-screened, versioned with rollback, and can be switched off.
Set organization-wide defaults and your own personal override to control when review missions auto-generate for new commits.
Understand
Pro and Team: one click shows what else in the codebase a change could affect, and how often similar past PRs later needed a fix or were reverted.
Pro and Team: an AI estimate of whether your existing tests would catch plausible mistakes in the changed code, naming the test and line that would catch each one. Nothing is executed; covers TypeScript/JavaScript, Java and Kotlin.
Pro and Team: a plain-English summary of every new or updated PR on your watched repos: what it does, a simple diagram and a review-effort estimate. Posting it to GitHub is off by default.
Per-tech-stack review checklists that owners and admins can customize, plus SonarQube violations scoped to exactly the files changed in the PR.
Pro and Team: post the AI-generated review to the PR yourself, when you choose. It costs no AI usage, and a stale review is refused if new commits have landed.
Connect
Choose your provider from the catalog and test the connection. On Pro and Team the key is registered with LiteLLM and never stored in Qodillo’s database.
Pro and Team: connect your own tool servers, such as Jira, Notion or an internal API, so the AI chat can look things up mid-review. Opt in per server, per tool and per repository; withheld on public repos unless you allow them.
Installs as a GitHub App with GitHub OAuth sign-in — no separate account, reviewing the repos you’ve already got.
Connect more than one GitHub identity to a single account instead of creating duplicates. An identity already linked elsewhere is blocked with a clear conflict, never silently merged.
Team
Owner, Admin, and Member roles share one token pool and one set of review standards. Invite up to 10 teammates; invites expire after 24 hours.
Team owners and admins can require members onto the managed AI pool or onto the team’s own AI key.
Paid plans: every grant and charge on record, plus a chart of which features use the most. Owners and admins see the whole organization, members see their own. Warnings at 80% and 95% of your allowance.
Paid plans: buy extra usage on top of your monthly allowance when you need it.
Pro and Team owners and admins can hide tech stacks you don’t use and turn off the “Reviewed with Qodillo” line on published reviews.
Operational workflow
GET /api/queue · application/x-ndjson
{"pr":"payments-api#412","risk":7.5}
{"pr":"auth-svc#56","risk":3.5}
{"pr":"web-app#1093","risk":3.1}
Risk-ranked queue · sample data
Security
PR diffs, commits, and CI data are fetched live from GitHub per request and never stored server-side — there’s no second copy of your code sitting in a database.
Free-tier LLM and SonarQube credentials live only in your own browser. On Pro and Team, your BYOK key is registered with LiteLLM and never stored in Qodillo’s database or shown again after saving — you choose which provider sees your code.
Qodillo connects as a GitHub App, not a personal access token — access is limited to exactly the repositories you install it on, and can be revoked from GitHub at any time.
Automated risk-based approval is off by default and never writes to GitHub until you turn it on — every decision is logged to a reviewable history, and any error during evaluation fails closed to requires_human_review instead of approving by accident.
Risk-scored pull requests, AI review missions for what to check first, and no source code stored.