Cross-model peer review for coding agents

Model Peer v0.8.1, an open-source tool from makedirectory, enables cross-model peer review for coding agents by letting Claude Code, OpenAI Codex CLI, and Google Gemini CLI consult each other as read-only peers, with the primary agent as the hub and reviewers running in parallel. The tool, installable via curl, creates per-project skills and commands, and it never reads or writes AGENTS.md, CLAUDE.md, or GEMINI.md.

Cross-model peer review for coding agents. Model Peer lets Claude Code, OpenAI Codex CLI, and Google Gemini CLI consult one another as independent, read-only engineering peers. model-peer ask codex "Review this authentication design for bypasses" model-peer review This is the part that surprises people. Your agent is the hub. Each consultation spawns another vendor's CLI, read-only, gets one answer, and exits. Nothing persists. php Primary agent | +-- independent peer model -- advisory response | +-- primary agent evaluates the advice The peer supplies evidence, not authority. Project rules and invariants still win. --depth deliberately relaxes this, and it is opt-in: php depth 1 default primary - peer - primary depth 1 opt-in primary - peer - peer - primary Even then the peer runs nothing. It asks Model Peer for the second opinion, and Model Peer decides whether to perform it. model-peer review keeps its reviewers as leaves unless you pass --depth , and even then a reviewer may only consult a model that is not on the panel. Reviewers that can consult each other are not independent observations, which is the whole point of the panel. The peer also starts in your working directory with read tools enabled, so you don't paste code into the question — name files and symbols and let it look. Coding agents can review their own work, but self-review is still self-review. model-peer review fans your Git diff out to every installed model independently. None of them sees the others' conclusions; only then does a synthesizer reconcile the findings. Because reviewers can't anchor on each other, agreement between them is real signal. Reviewers are independent, so they also run in parallel: a review costs roughly the slowest model rather than the sum of them, and Model Peer waits for the whole panel before synthesizing. php +-- Claude --+ | | git changes --+-- Codex ---+-- synthesis | | +-- Gemini --+ curl -fsSL https://raw.githubusercontent.com/makedirectory/ModelPeer/v0.8.1/install.sh | bash Or clone and run ./install.sh . As with any remote shell installer, inspect it first. Model Peer never asks you to paste an API key — authentication stays with each vendor CLI. Installing Model Peer globally gives you a command. It does not give the coding agent in your repository a habit. One command per project fixes that: cd ~/code/your-project model-peer init all or: init claude / init codex / init gemini Name the agents whose directories it may write. There is no default — a repository gets the CLIs its team actually uses, and nothing else. created .claude/skills/cross-model-review/SKILL.md created .claude/skills/cross-model-consult/SKILL.md created .codex/skills/... both, per CLI created .gemini/skills/... created .claude/commands/peer-review.md created .claude/commands/peer-ask.md Two skills, because the tool does two things that fire on different cues: review cross-checks a diff across the whole panel, consult gets one peer's opinion on one question. Each is a self-contained directory in the place each vendor set aside for skills. Your AGENTS.md, CLAUDE.md, and GEMINI.md are never read, written, or symlinked. Now the agent consults a peer on its own — before an architecture decision, on a bug that has outlived two hypotheses, on anything security-sensitive — and tells you which model it asked and whether it took the advice. In Claude Code you also get /peer-review for a full cross-model review of the current diff. Commit those files and your team gets the same behavior. After upgrading Model Peer, model-peer update refreshes them; model-peer update --check verifies them in CI. → In your workflow https://modelpeer.app/workflow · Agent skills https://modelpeer.app/agent-rules model-peer ask <claude|codex|gemini "<focused question " consult one peer model-peer review "focus instructions" cross-model review model-peer init <claude|codex|gemini|all install the skills model-peer update --check refresh them model-peer trust let Gemini load them model-peer doctor --probe check setup Every consultation is bounded --timeout , 600s default and reports progress on stderr. A reviewer that hangs, fails, or returns nothing is dropped and named rather than taking the whole panel with it. → Usage https://modelpeer.app/usage · CLI reference https://modelpeer.app/reference Peers are launched with the most conservative non-interactive configuration each vendor supports: Plan mode or a read-only sandbox, no file-editing tools, no general shell, and stdin closed. Model Peer stores no credentials. Where a vendor reads trust settings from the environment, Model Peer supplies them itself rather than inheriting yours. Gemini's GEMINI CLI TRUST WORKSPACE is cleared before launch: it would otherwise enable MCP servers declared by the workspace, which Gemini starts as local subprocesses during tool discovery — before any tool policy applies. A reviewed repository does not get to configure its own reviewer. Consultation chains are bounded by --depth default 1, ceiling 10 , and a model is never consulted twice in one chain. Depth is a limit, not a permission — raising it increases how many models can participate, never what a model can do to your system. A peer that wants a second opinion asks for one in its reply; Model Peer validates the request and performs the consultation itself. No peer is given a shell, at any depth, and no peer can start a consultation Model Peer did not authorise. --timeout is one deadline for the whole invocation, not a fresh budget per hop, so --depth 3 --timeout 600 is a ten-minute operation rather than a possible fifty. Under review the deadline is per reviewer, and a consultation a reviewer starts comes out of its own budget. Reviews cover untracked files as well as tracked changes, so new code is reviewed rather than merely listed. This is defense in depth, not a formal sandbox. Upstream CLI behavior can change — model-peer doctor --probe runs one real consultation per CLI and verifies on disk that none of them wrote anything, which is the only way to know it still holds after a vendor upgrade. → Safety boundaries https://modelpeer.app/safety · Peer-chain depth https://modelpeer.app/depth · Troubleshooting https://modelpeer.app/troubleshooting make test smoke tests against stub CLIs; no model usage make lint syntax check, plus shellcheck when installed make sync regenerate install.sh's embedded copy of bin/model-peer → Development https://modelpeer.app/development · Roadmap https://modelpeer.app/roadmap Questions and ideas are welcome in Discussions https://github.com/makedirectory/ModelPeer/discussions . Security issues should go through private reporting https://github.com/makedirectory/ModelPeer/security rather than a public issue. If a peer caught something before it shipped, I'd genuinely love to hear about it. If a peer told you something confidently wrong — we've never met, and this is the first you're hearing of it. Peers are advisory. It says so above. Either way, if it saved you a review cycle, you can buy me a coffee: MIT, Copyright c 2026 Make Directory Developers, LLC. See LICENSE /makedirectory/ModelPeer/blob/main/LICENSE . Model Peer is an independent open-source project maintained by Make Directory Developers, LLC. It is not affiliated with, endorsed by, or sponsored by Anthropic, OpenAI, or Google. Claude, Codex, and Gemini are trademarks of their respective owners.