Open-source CLI · local-first · MCP server
Never write the same review comment twice.
difflore turns past PR reviews and your mid-session corrections into rules your agents follow — fewer repeat comments, less rework, shorter reviews.
Open-source runtime
Start locally with the CLI and MCP server. No account needed.
CI gate
Check every pull request against the same approved rules, with your own model key.
Apache-2.0 · runs locally · no account · works with 12 AI coding tools
How it works
A small approval loop, not another knowledge base.
difflore stays useful by keeping the rule set small, source-linked, and approved by the people who own the code.
- 1
Import review history, capture live corrections
Start with merged PR comments where your team already said what mattered; the corrections you make in live agent sessions keep new candidates coming.
- 2
Approve reusable rules
Accept only patterns worth repeating; dismiss one-off comments before agents see them.
- 3
Agents receive rules
Claude, Codex, Cursor, and MCP clients receive the approved rule before they touch similar code.
Injection guardrails you can verify in the open-source runtime
- Every injection fits a ~1,500-token hard budget.
- Injects after each edit, with a 120s per-file dedup window — and goes quiet automatically when recall comes back empty.
- Strict per-repo isolation: rules learned in one repo are never recalled in another.
- Every injection is logged with its token cost; every citation an agent makes is fed back as telemetry.
Start small
Try it on one repo before changing your workflow.
Import real review history, approve the obvious wins, and expand once the rules keep proving useful.
Imports from
Rules show up where your AI coding assistant already runs.
- GitHub PR reviews
Agents
- Claude Code
- OpenAI Codex
- Cursor
- Windsurf
- Gemini CLI
- Goose
Launch path
- Start with the free local CLI
- Import one repo of GitHub review history
- Add the rule check to CI once one repo is working
Common questions
Accepted PR review feedback plus the corrections you make in live coding sessions, the rules your team approved, and the evidence that explains where each one came from.
No. New candidates wait for approval, so noisy or one-off comments can be dismissed before agents ever see them.
Claude Code, Codex, Cursor, Gemini CLI, Windsurf, Goose, and the rest of the 12 supported AI coding tools — any agent that can read MCP or HTTPS context can connect.
Not to us. Rules and activity live in a local SQLite file. Review import and rule drafting run through the agent CLI or API key you already use, so your code only reaches the model provider you picked.
Committed context files go stale and carry no provenance. difflore treats CLAUDE.md, .cursor/rules, and AGENTS.md as export surfaces; underneath, the system of record keeps every rule human-approved, traceable to the review comment it came from, and triggered only on the paths where it applies.