Config-driven
Add a new reviewer by editing config.json. No code changes, no
per-provider classes, no plugin system.
llm-review-panel
One plan, several LLMs, one consolidated review. Locally, no API keys.
What it does
You write a markdown plan (an implementation plan, a review request, an audit
prompt). llm-review-panel sends that same plan to multiple LLM command-line
tools running on your machine, captures each review independently, then asks
one of them to synthesize the lot into a single consolidated review that
tags findings as consensus, contested, or singleton.
The point is to get past single-model bias. One LLM agreeing with itself is not a review. Three different model families agreeing is a signal.
No API keys
Reviewers are local CLIs you already have installed and authenticated (Claude Code under your subscription, OpenCode against local Ollama, Gemini CLI on its free tier, etc.).
Parallel
Reviewers run concurrently up to max_parallel. Wall-clock stays close
to the slowest single reviewer.
Human in the loop
Three checkpoints: confirm the prompt, inspect raw reviews, accept the
synthesis. --yes skips them; --dry-run validates without spawning.