What it is

Pick a branch. Pick a base. Hit ⌘⌥R.

Claude reviews the diff five times — exploring it, then through a security lens, a performance lens, an accessibility lens, a tests lens — and then critiques its own findings to drop the noise. Every comment lands on an exact file:start-end, with reasoning, evidence, alternatives and a suggested fix you can preview before applying.

It runs entirely on your Claude Code CLI. No API key to manage, no separate billing.

Think of it as the reviewer who reads the whole PR, asks "what could go wrong here?" on every chunk, and then has the self-awareness to retract their own bad takes before you see them.

Why you might want this

• You ship faster than reviewers can keep up. Get a first pass in 5 minutes, locally, before opening the PR.
• You've been burned by AI tools that hallucinate findings. Every finding here is anchored to real lines, with quoted evidence, and survives a self-critique pass.
• You don't want another API key. If claude works in your terminal, this works.
• You want the receipts. The panel shows tokens, cache hit rate, which passes ran, and what critique dropped — no black box.

Install

From VS Code — search "Claude Branch Reviewer" in Extensions, or open the Marketplace page.

From the terminal

code --install-extension ajmc90.claude-branch-reviewer

You'll also need the Claude Code CLI installed and logged in. Run claude once in your terminal — if that works, you're done.

Quick start

1. Open a git repo in VS Code.
2. ⌘⇧P → Claude Review: Open Review Panel (or just press ⌘⌥R / Ctrl+Alt+R).
3. Pick a base and head branch. The panel shows ahead/behind, staleness, age and author.
4. (Optional) Toggle which passes to run, or pick a preset: fast · deep · security · performance · accessibility.
5. Hit Start review. Wait ~3–6 minutes for a typical PR.

Findings stream in live. Each one has a one-click jump-to-code, a diff-preview Apply Fix, and a follow-up button that opens a Claude terminal preloaded with the context.

How it works

A review runs five phases, in order. The last one cleans up after the first four.

A · Discovery        →  Reads the diff. Builds a change map. Surveys hot spots.
B · Specialists      →  Security · Performance · A11y · Tests   (in parallel)
C · Consolidation    →  Dedupe + cluster. No CLI call — pure local logic.
D · Completeness     →  Gaps Claude expected but didn't find. Trade-off analysis.
E · Self-critique    →  Re-reads every finding. Keeps, revises, drops, or merges.
                        Emits the final verdict + executive summary.

The critique phase is the one most AI review tools skip. It's the one that catches:

• Findings that sounded smart but don't survive a re-read.
• Three specialists flagging the same bug from different angles → merged into one.
• Severity inflation → bumped down with a recorded reason.

Nothing disappears silently. Dropped and merged findings live under a Revised tab with the critique's reasoning, so you can audit what it threw out.

→ Full pipeline details

What you actually see

While it runs. The right pane becomes a stage — phase eyebrow (PHASE 3 OF 5), elapsed timer, four-segment progress bar, live signals (tokens · files · passes · discoveries), and a per-pass timeline. Findings stay hidden until critique finishes, so you don't watch findings appear and then vanish.

When it finishes. A verdict, risk score, executive summary, and a filterable findings grid: Problem ↔ Solution cards with severity ribbons, category badges, jump-to-code, apply fix, follow-up, and a per-card EN/ES translation chip.

Always visible. The activity-bar sidebar mirrors state (idle / running / paused / failed / done), shows the last review's verdict, and keeps a history of the last 5 reviews per (base, head) pair. One click rehydrates.

→ Detailed UI tour

Cost, honestly

Before you press RUN, a cost pill estimates the run for this diff with these passes:

~95K tokens · ~6 min   [calibrated]

Click it for a breakdown by bucket (Fresh input · Cache creation · Cache reuse · Output) and per pass. After 5+ runs in your workspace, the estimator self-calibrates to your repo + machine.

No dollar figure. You're on a Claude subscription; you don't pay per token. Surfacing dollars would lie about your bill. Tokens + cache reuse % are the honest unit.

A confirmation modal only fires for runs over ~3M effective tokens — the threshold where one review meaningfully bites into a 5-hour subscription window. Anything smaller, just go.

→ How the estimator works

Features worth knowing about

• 🌿 Branch freshness gate — flags stale branches before review starts. One-click ff-only pull for base + head (without checking out either). Sync head with base via Merge or Rebase, all from the banner.
• ♻ Phase-scoped session reuse — passes inside the same phase share a Claude CLI session, dropping effective tokens ~65% vs. running passes in isolation.
• 🧠 Self-critique audit trail — every prior finding tagged keep / revise / drop / merge with a reason. Nothing vanishes silently.
• 🔁 Apply Fix preview — fixes open as a VS Code diff editor. Edit the right side, then Apply or Discard from the title bar. Claude's suggestion is a starting point, not the final word.
• ⏸ Pause, resume, retry — failed or cancelled reviews snapshot themselves. Resume picks up where it stopped; per-pass Retry for the one that flaked.
• 🔕 Silence noise — dismiss findings as "this one" or "this pattern, everywhere". Future matches return muted with a 🔕 badge, restorable any time.
• 🌐 Bilingual UI — English/Spanish UI with on-demand per-finding translation (EN/ES chip on each card).
• 🧩 Adapts to your project — auto-detects language, framework and tests; reads CLAUDE.md, README.md, CONTRIBUTING.md, ARCHITECTURE.md.

FAQ

Requirements

• VS Code 1.85 or newer
• A git repository open as your workspace
• Claude Code CLI installed and logged in — claude once in a terminal walks you through it

Privacy

The extension does not send code to any third-party API of its own. It shells out to the claude CLI on your machine, which talks to Anthropic under your account. Secret files (.env*) are stripped from the prompt by default — the review sees that they changed, not their contents. All state is stored in VS Code's workspace storage. No telemetry.

Docs

• 📖 Pipeline & passes — what each phase does and why
• ⚙️ Settings — every config key, what it does, what to change
• ⌨️ Commands — every command and keybinding
• 💰 Cost & calibration — how the estimator works
• 🏗️ Architecture — for contributors

Develop locally

This project uses pnpm.

pnpm install
pnpm run compile
# Press F5 in VS Code to launch a dev host with the extension loaded

See docs/ARCHITECTURE.md for the source tree, and RELEASE.md for the publish flow.

Support

If this caught a bug before your reviewer did, or saved you an hour of "what is this PR even doing", consider buying me a coffee.

⭐ Stars on GitHub and reviews on the Marketplace also help — they make this project visible to other devs.

License

MIT — see LICENSE.