Local AI-powered code review agents for Claude Code
We believe code reviews should be faster and more efficient, focused on trade-offs, knowledge sharing and accountability, not on policing style, naming quirks, or catching trivial regressions and security issues.
Use AgentCheck to run specialized agents before you commit, directly in your dev flow. AgentCheck pulls in your real project context (docs, conventions, architecture) and local ecosystem (MCP tools, resources) to deliver focused, actionable feedback.
- ✅ Local review — catch issues while you’re coding, not after PRs
- 🧠 Project-aware — understands your docs, conventions, and methodology
- 🔧 Open source & customizable — adapt each agent to your stack and standards
- 🖥️ Integrated flow — runs inside your environment, no black-box platforms
- 💸 Free to use — no seat licenses, no subscription fees
# In your project root
git clone https://github.com/devlyai/AgentCheck.git .claude-temp
cp -r .claude-temp/agentcheck .claude/
cp -r .claude-temp/agents .claude/
rm -rf .claude-temp
echo ".claude/agentcheck/reports/" >> .gitignoreThis adds:
.claude/
agentcheck/ # context and outputs
agents/ # the review agents
Ask Claude Code to analyze your project and customize the agent templates:
"Initialize AgentCheck context for this project"
This analyzes your codebase and generates project-specific context files in .claude/agentcheck/project-context/.
Add to your AGENTS.md / CLAUDE.md:
## AgentCheck — Local AI Code Review
Before committing ask Claude Code to "Run AgentCheck analysis"
The agents (logic-reviewer, security-checker, guidelines-compliance,
style-compliance, product-reviewer) analyze **staged changes** and write reports to:
`.claude/agentcheck/reports/`.- Diff‑aware reviews focused on what you changed
- Prioritized findings: 🔴 Critical · 🟠 High · 🟡 Medium · 🔵 Low
- Concrete fixes: suggestions you can apply immediately
- Project‑specific context: agents learn from your docs and past decisions
All agents running in parallel:
Comprehensive analysis results:
| Agent | Focus | Typical Findings |
|---|---|---|
| logic-reviewer | Correctness, edge cases, regressions | Broken invariants, unsafe refactors, missing tests |
| security-checker | Vulnerabilities & secrets | Injections, insecure defaults, hardcoded creds, dangerous deps |
| guidelines-compliance | Team standards & methodology | Violations of ADRs, naming/domain rules, missing docs |
| style-compliance | Consistency & readability | Lint parity, formatting drifts, dead code, unclear APIs |
| product-reviewer | User impact & behavior changes | UX regressions, backward compatibility, telemetry/analytics gaps |
Reports are written to .claude/agentcheck/reports/ and grouped by agent.
- Collects context: project docs, ADRs, coding standards, and your Claude Code MCPs
- Analyzes staged diffs per agent specialization
- Ranks issues and proposes concrete, minimal‑diff fixes
- Continuously adapts by generating and refining project‑specific context files
AgentCheck is designed to get quieter and smarter over time.
- Edit agent prompts in
.claude/agents/*to encode team standards and false‑positive rules - Add project context: link to docs/architecture, ADRs, and service READMEs the agents should prefer
- Leverage MCPs: your Claude Code MCP tools (project management, knowledge management, local resources, context manager) give agents deeper context
- Team calibration: as prompts evolve, human CR comments will have higher focus on design and be more impactful.
Goal: human code reviews faster and more useful — not a daily frustration.
- No reports generated → Ensure your agent has relevant permissions and paths configured correctly.
- Too many false positives → Tighten prompts in
.claude/agents/*and add links to your canonical docs. - Agent misses context → Point agents to the right docs/ADRs; ensure your MCPs are available in Claude Code.
How is this different from PR bots (e.g. Cursor Bugbot)? PR bots only run after code is pushed, so feedback arrives late, outside the developer’s flow. They’re closed systems, often noisy with trivial comments, and charge per seat. They try to solve problems too late, in the wrong place, and without your developer environment tools.
Does it block commits? No. AgentCheck doesn’t block anything by default; you choose when to run it before committing.
PRs and issues welcome! If you extend an agent (new checks, better prompts, project context helpers), include a short rationale and examples.
MIT — adapt and evolve for your team’s needs.
Attribution: AgentCheck is not affiliated with Anthropic. “Claude” and “Claude Code” are trademarks of their respective owners.