Multi-agent AI code reviewer for VS Code.
Five specialized agents. Evidence-backed findings. Validated patches. Full audit trail.
Finding detail — evidence, code context, Generate Fix and Verify Fix pipeline
Live findings list — Fixed / Still Present badges, severity, confidence scores
Audit timeline — every agent action, tool call, and finding timestamped
Most AI code reviewers are single-agent, text-output-only, and non-replayable. ChainReview is different.
It runs five specialized agents in parallel inside VS Code — each grounding its analysis in real static tooling (Semgrep, TypeScript compiler, ripgrep, import graph). Findings come with file paths, line ranges, code snippets, and confidence scores. Patches are syntax-validated before you see them. Every agent action is recorded in a local SQLite audit log.
Built for the Built with Claude: Claude Code Hackathon by Anthropic.
| 5 specialized agents | Architecture · Security · Bugs · Validator · Explainer |
| Evidence-backed findings | File path + line range + code snippet + confidence score |
| Validated patches | TypeScript compiler + clean-apply check before you see a fix |
| Coding agent handoff | Send findings to Claude Code, Cursor, Windsurf, Copilot, or Codex |
| Audit trail | 8 event types recorded in local SQLite — fully replayable |
| @mention routing | @security, @bugs, @architecture — run only what you need |
| Diff mode | Review only staged/unstaged changes — fast PR-style feedback |
| MCP extensible | Add external MCP servers from the UI |
| Agent | What it catches |
|---|---|
| Architecture | Coupling, circular dependencies, boundary violations |
| Security | Injection, auth gaps, crypto issues, data exposure |
| Bugs | Logic errors, null refs, race conditions, edge cases |
| Validator | Challenges findings from other agents — cuts false positives |
| Explainer | Deep-dive on any finding, on demand |
Architecture, Security, and Bugs run in parallel via Promise.allSettled. Validator then challenges the combined result. Explainer fires on demand.
Prerequisites: Node.js 18+, VS Code 1.90+, Anthropic API key, Semgrep (optional: brew install semgrep)
git clone https://github.com/charannyk06/ChainReview.git
cd ChainReview
npm install && npm run buildPress F5 in VS Code → Extension Development Host → click the </> ChainReview icon in the activity bar.
# Set your API key in VS Code Settings > ChainReview, or:
export ANTHROPIC_API_KEY="sk-ant-..."Finding card
├── Explain → Explainer agent deep-dive
├── Generate Fix → LLM-generated unified diff patch
├── Verify Fix → Validator re-checks the proposed fix
├── Handoff To → Claude Code / Cursor / Windsurf / Copilot / Codex
└── False Positive → Dismiss and record
Patches go through TypeScript syntax check (ts-morph) and clean-apply verification before you see them. "Fixed" and "Still Present" badges update after Verify.
| Mode | What gets reviewed |
|---|---|
| Full Repository | Entire codebase — file tree, import graph, Semgrep |
| Diff Review | Staged + unstaged changes only — fast PR-style feedback |
| Chat | Ask questions about the codebase using the same CRP tools |
Eight event types, timestamped and stored locally in SQLite (~/.chainreview/chainreview.db):
agent_started → evidence_collected → finding_emitted → patch_proposed
→ patch_validated → human_accepted → human_rejected → false_positive_marked
Browse the full timeline in the Audit Trail tab. No telemetry. No cloud sync. Fully replayable.
ChainReview is designed with a local-first, defense-in-depth security model. No code or review data leaves your machine unless you explicitly initiate a request to the Claude API.
- Local-first storage — all review data is stored in
~/.chainreview/chainreview.db(SQLite). No telemetry, no cloud sync. - Local static analysis — Semgrep runs entirely on your machine. No code is sent to external scanning services.
- Secrets redaction — API keys, tokens, and passwords are stripped from code snippets before they are sent to the Claude API.
- No destructive actions without confirmation — patches require explicit manual Apply. No file is modified automatically.
- Path traversal protection — the patch engine enforces repository boundary checks. Patches cannot write outside the root of the open repository.
- Allowlisted shell execution — the
crp.exec.commandtool only runs commands from an explicit allowlist. Arbitrary shell execution is blocked. - API key handling — your Anthropic API key is stored using VS Code's built-in secret storage and is never written to disk in plaintext.
- Minimal network surface — the only outbound connections are to
api.anthropic.com(LLM calls) and, optionally, the Brave Search API if web search is enabled.
ChainReview Protocol (CRP) is an open, MCP-compatible tool schema. All tools are available to agents and the chat interface.
Full tool reference (20+ tools)
| Tool | Description |
|---|---|
crp.repo.open |
Initialize a repository for review |
crp.repo.tree |
File tree with depth and pattern filtering |
crp.repo.file |
Read file contents with optional line range |
crp.repo.search |
Ripgrep search with regex, glob, result limits |
crp.repo.diff |
Git diff (staged, unstaged, or between refs) |
| Tool | Description |
|---|---|
crp.code.import_graph |
TypeScript import graph with cycle detection (ts-morph) |
crp.code.pattern_scan |
Semgrep static analysis with configurable rules |
| Tool | Description |
|---|---|
crp.patch.propose |
Generate a unified diff patch for a finding |
crp.patch.validate |
Validate patch (clean apply + TypeScript syntax) |
crp.patch.apply |
Apply a validated patch to disk |
crp.patch.generate |
Generate fixed code via LLM reasoning |
| Tool | Description |
|---|---|
crp.review.run |
Run full review pipeline |
crp.review.cancel |
Cancel an active review |
crp.review.get_findings |
Query findings by run ID |
crp.review.get_events |
Query audit events by run ID |
crp.review.validate_finding |
Run validator on a single finding |
crp.review.record_event |
Record a chain-of-review event |
| Tool | Description |
|---|---|
crp.chat.query |
Ask a question about the repository |
crp.exec.command |
Run allowlisted shell commands |
crp.web.search |
Web search via Brave Search API |
| Layer | Technology |
|---|---|
| Extension | VS Code Extension API, esbuild |
| Server | Node.js, MCP SDK, better-sqlite3, ts-morph, simple-git |
| UI | React 18, Framer Motion, Lucide Icons |
| LLM | Anthropic Claude API (claude-opus-4-6) |
| Static Analysis | Semgrep (local) |
| Build | Vite 6, esbuild, TypeScript 5.7 |
- VS Code Marketplace publish
- CLI tool for CI/CD integration
- GitHub PR commenting + draft PR creation
- Azure DevOps native integration
- Multi-language support beyond TypeScript
- Custom agent creation
npm install
npm run build # all targets
npm run dev # watch mode
npm run test # all testsPowered by Claude · Model Context Protocol · Semgrep · ts-morph