Gauntlet
Gauntlet is a read-only MCP server that sends Claude's code, diffs, plans and questions to other model families (Codex, Gemini, GPT-OSS) for second-opinion reviews, diagnosis, analysis and audits.
quick_check — fast, cheap second look at a small item: a function, regex, query, config value or one-file change.
codex_review / gemini_review — adversarial or contract/state-focused review of a diff or a few files, returning a verdict and file:line findings checked against the files.
codex_diagnose — root-cause analysis of a bug or failing test from a symptom, error output and a few files.
gemini_analyze — explains how a broad or unfamiliar part of the code fits together before you edit it.
security_audit — finds exploitable issues: injection, authz gaps, secrets, SSRF, path traversal, races, client-input trust.
edge_cases — surfaces mishandled inputs and states as concrete test cases.
plan_critique — critiques an implementation plan before code is written (wrong assumptions, missing steps, simpler paths, risks).
council — asks Codex, Gemini and GPT-OSS the same question in parallel (review, diagnose or critique mode) and merges verdicts, agreement findings and disagreements.
gauntlet_status — reports which CLIs are installed and signed in, the model per tier, paused models and limits in force.
Common controls across tools: pick a
tier(light/standard/deep/auto) or pin amodel, scope withpaths,checks,constraints,difforplan, settimeoutMs, and skip the cache withnoCache. All tools are read-only and write nothing.
Provides Gemini-based code analysis and second-opinion reviews, focusing on interface contracts, error paths, state, and user-facing regressions. Gemini can also participate in council parallel reviews alongside other model families.
Provides adversarial code review, root-cause diagnosis, plan critique, edge-case generation, and security auditing via OpenAI Codex and GPT-OSS models. The server sends objectives, diffs, files, and constraints to these models and returns structured verdicts and file-checked findings.
Click on "Deploy Server".
Wait a few minutes for the server to deploy. Once ready, it will show a "Started" state.
In the chat, type
@followed by the MCP server name and your instructions, e.g., "@Gauntlet review my staged changes and flag any real bugs"
That's it! The server will respond to your query, and you can continue using it as needed.
Here is a step-by-step guide with screenshots.
Gauntlet
Make Claude's code run the gauntlet.
Codex, Gemini and GPT-OSS review what Claude Code writes. Every finding has to quote the line it's about, and Gauntlet checks that quote against your files before Claude reads it. It runs on the ChatGPT and Google plans you already pay for. No API keys.
In 157 real reviews of code Claude Opus had just written, another model family found a serious issue 62% of the time, and Claude went on to change the flagged file in 95% of those cases. How that was measured
Why I built this
I do most of my coding with Claude Code, and I kept hitting the same two walls.
The first: asking Claude to review its own code mostly gets you Claude agreeing with itself. The bug it didn't see while writing, it doesn't see while reviewing.
The second came when I started asking other models instead. They found real bugs, and they also made some up. A confident paragraph about a race condition on line 84, and line 84 is a comment. I'd lose half an hour finding that out.
Gauntlet is what I ended up with. Claude still writes the code and is still the only thing allowed to touch my files. When a change matters, it sends a small packet (the objective, the files, the diff) to a model from a different family and gets a structured review back. Before Claude sees that review, Gauntlet looks up every quoted line in the real file and sorts the findings into three groups:
verified;
nothing to check against a file;
"could not be confirmed", meaning the quoted code isn't there.
Here's a real run, Gauntlet reviewing its own JSON parser with two families:
# Council (review): 2/2 voices answered
## Verdicts
- codex (gpt-5.6-terra): approve, confidence high
- gemini (gemini-3.8-flash-high): approve_with_changes, confidence high
Disagreement: approve vs approve_with_changes. Resolve it with evidence, not by averaging.
## gemini
[gemini / gemini-3.8-flash-high, advisory and read-only: tier standard, 108.4s,
2/2 claims verified against the files]
## Findings verified against the files (2)
1. [medium] src/json.mjs:32: extractJson aborts on the first unclosed opening brace,
failing to extract subsequent valid JSON objects.
Fix: Replace `break;` with `continue;` ...
2. [low] src/json.mjs:28: extractJson returns arrays when the input is a standalone
JSON array ...Codex approved the file. Gemini found two real bugs, both quoted from the source, and both got fixed in the next commit. That's the whole idea in one screen.
Related MCP server: claude-reviewer-for-codex
What's in it
An MCP server with ten tools: reviews, a diagnosis tool, plan critique, a security audit, edge cases, and
council, which asks Codex, Gemini and GPT-OSS the same question at once and tells you where they agree. Findings that two families reached on their own are the ones I trust most.A fallback chain, because quotas run out. If the flagship model is rate-limited, the call goes to the smaller model of the same family, then to another family, then to Claude through Antigravity. A model that just failed sits out for a while, so the next call doesn't wait on it again.
Four read-only Claude subagents (
haiku-navigatoron Claude Haiku 5.5,sonnet-reviewer,opus-architect, andsonnet-test-analyst, which is opt-in) for when you'd rather stay in the Claude family. See subagent models.Optional hooks:
a pre-turn hook that runs a small pipeline of internal agents (comprehension, anti-hallucination, text review, jury, legal, learning);
a gate that refuses
git pushwhile edited code hasn't had a review.
The pipeline config is versioned. Publish a new version and chats that are already open pick it up on their next message, without a restart. It's the part I'm proudest of, and the one I'd point a curious engineer at first. See architecture.
Does it actually help?
I use Claude Opus every day, and I wanted to know whether the second opinions were worth the wait. There are two ways to measure it, and the answers differ in an interesting way.
On real work
Claude Code keeps every session on disk. bench/real-world.mjs
reads those transcripts and, for each review, checks three things:
was the reviewed code Claude's own, written earlier in that same session;
did the review raise a serious issue (critical, high or medium);
what did Claude do next: edit a file the finding named, or call it a false positive.
I ran it on three weeks of my own sessions on two machines: 73 sessions, Claude Opus 5 and 5.5, real projects.
In 157 reviews of code Claude had just written and was ready to move on from, another model family found a serious issue 62% of the time. Claude then edited the file the finding named in 95% of those cases, and dismissed 2 of 98 as false positives.
That matches what using it feels like. I've almost stopped reporting bugs to Claude myself. What's left are misunderstandings about what I asked for, not mistakes in the code.
The method has limits, and you should know them:
"Edited the flagged file" means Claude changed a file the finding named, after the review and before the next one. That's a strong sign the finding was accepted, but it doesn't prove each edit fixed exactly that finding.
The data is from one developer's sessions.
The script is in the repo, so run it on your own history and see what you get:
node bench/real-world.mjs ~/.claude/projectsA controlled benchmark
I also wanted a test anyone can rerun, so I planted bugs in 16 code samples and gave the same files to four reviewers, three times each: 192 reviews.
Twelve samples have one bug I put there on purpose, and they're not textbook ones:
an LRU cache that never refreshes recency;
a semaphore that leaks a permit on cache hits;
cents passed to an API that expects euros, across two files;
a token bucket that refills in milliseconds instead of seconds;
an Express route registered before its auth middleware;
a booking race;
a ReDoS-prone regex;
and five more.
The other four files are correct. Three model families checked them and found nothing, and they're there to count false alarms. A bug only counts as found when the finding points at the right lines and names the actual problem. The method, the corpus and the raw answers are in docs/benchmark.md.
What the controlled test shows, and why it differs from the real sessions:
Here Claude found every planted bug on its own. These bugs were written by me, in short self-contained files, and Claude was asked to look for them. That is the easy case. On real work, Claude is reviewing code it wrote itself, minutes earlier, in the middle of a long task, convinced it's right. That's where the 62% above comes from, and no synthetic test I can build reproduces it well.
The council cuts the noise. Claude alone raised 2 false alarms per run on code with nothing wrong in it. Keep only what at least two of four families agree on, and you still get 12 of 12 bugs with 0.7 false alarms per run.
Each reviewer has fixed blind spots. Codex missed the ReDoS in all three runs, and GPT-OSS missed the mutable default argument every time. A different family covers them.
Made-up findings are rare but real. 2-8% of the findings from Claude, Codex and GPT-OSS quoted code that isn't in the file. Gemini had none. Gauntlet looks up every quote and labels those before Claude reads them:
Reviewer | Bugs found (of 12) | False alarms / run | Median time per review |
Claude Sonnet 5.5 (alone) | 12 | 2.0 | 12 s |
Codex GPT-5.6 Terra | 11 | 0.3 | 10 s |
Gemini 3.8 Flash | 11.7 | 0.7 | 73 s |
GPT-OSS 120B | 8.7 | 2.0 | 86 s |
Gauntlet council (any family) | 12 | 3.0 | 107 s |
Gauntlet council (2+ families agree) | 10.7 | 0.3 | 107 s |
Claude + Gauntlet council (2+ of 4 agree) | 12 | 0.7 | 107 s |
Some caveats:
Sixteen samples is small.
I wrote the bugs.
The models change week to week.
The council's time is its slowest voice, because the families run in parallel.
GPT-OSS didn't answer 6 of its 48 reviews; those count as misses, not as somebody else's answer.
There's also an easy set of textbook bugs, where every frontier model scores 12 of 12. That's why the set above exists.
A word on tokens. Gauntlet doesn't make reviews free. It moves them off your Claude plan and keeps the packets small: a median review above cost between 9k and 26k tokens on the other provider's side, depending on the model, and none on Claude's. docs/token-savings.md explains where the savings come from and how to measure them on your own work.
Installing
1. What you need
Node.js 20 or newer (
node --version).Claude Code, with the
claudecommand working in your terminal.At least one of the two provider CLIs below. You don't need both; Gauntlet uses whatever is installed.
2. The provider CLIs
Codex, using your ChatGPT plan:
npm install -g @openai/codex
codex login # sign in with your ChatGPT accountAntigravity, using your Google account. This is the one that gives you Gemini, GPT-OSS
and the Claude fallback lane. Install Google's Antigravity CLI (agy) by following
Google's instructions for your OS, then run it once so it can sign you in:
agy3. Gauntlet itself
git clone https://github.com/Raffymimii/gauntlet.git
cd gauntlet
npm install
node bin/gauntlet.mjs init --dry-run # prints every change, touches nothing
node bin/gauntlet.mjs initWhat init does by default:
registers the MCP server with
claude mcp add --scope user gauntlet ...;copies the read-only subagents into
~/.claude/agents.
Without a flag it doesn't touch your settings.json or your CLAUDE.md. These are the
opt-ins:
node bin/gauntlet.mjs init --claude-md # adds one @-include line to ~/.claude/CLAUDE.md
node bin/gauntlet.mjs init --hooks turn # pre-turn hook: live config + internal agents
node bin/gauntlet.mjs init --hooks review-gate
node bin/gauntlet.mjs init --test-analyst # the subagent that runs your testsI'd recommend --claude-md. It's the file that tells Claude when a quick check is enough
and when to call the council. Without it, Claude only uses the tools when you ask.
If you want a plain gauntlet command instead of node bin/gauntlet.mjs, run npm link
inside the folder.
4. Check it
node bin/gauntlet.mjs statusYou should see your CLIs as installed and signed in, the model for each tier, and no paused models. Then open a new Claude Code chat (MCP servers are only loaded when a chat starts) and try:
get a quick_check of src/whatever.ts: is the retry loop bounded?
run a council on this diff before I push it
5. Removing it
node bin/gauntlet.mjs uninstallThis removes the MCP server, the subagents it installed, its hook entries and its
CLAUDE.md line, and nothing else. Your data in ~/.gauntlet stays until you delete it.
When something doesn't work
Symptom | Likely cause |
Claude doesn't see the | You're in a chat opened before |
| Run |
"model not available" errors | Providers rename models. Put the current names in |
A review times out | The packet is too big, or the tier too high. Name fewer files, or ask for |
A model is "paused" | It failed on quota or login recently. It comes back by itself. |
| Set |
Safety
The reviewers are read-only:
Codex runs in its read-only sandbox, and Antigravity in plan mode with its sandbox on;
both have a hard timeout and a scrubbed environment;
the prompt goes in on stdin, never on the command line;
a reviewer can't call another reviewer.
What leaves your machine is the packet, and only the files you name. Gauntlet refuses:
.envfiles and keys;anything under
.ssh,.awsor.git;symlinks that lead outside the project;
binaries.
It also redacts token-shaped strings. gauntlet packet shows you exactly what would be
sent, without sending it.
What you send to OpenAI or Google is covered by your own agreement with them, and checking that your use fits your plans' terms is on you. Gauntlet uses the official CLIs under your account and backs off when a provider says you're out of quota.
The full picture is in docs/security.md, including what it doesn't protect against.
FAQ
Does it cost anything per token? No. It uses your existing logins through the official CLIs.
Can a reviewer change my code? No. Reviewers only answer; Claude decides what to apply.
Why not just ask Claude to review its own work? Sometimes that's fine. But another family catches different things. And a reviewer whose quotes get checked against the file can't send you after a bug that doesn't exist.
Will this work with my model names? The defaults are what I use. When yours differ,
~/.gauntlet/config.json overrides any of them.
Contributing
Issues and PRs welcome. If you touch a prompt in prompts/, run the agent regressions
(node bin/gauntlet.mjs regress) and paste the result in the PR. If you measure the token
savings on real work, I'd love to see the numbers, along with how you measured them.
License
MIT
Available Tools
10 toolscodex_diagnosecodex_diagnoseA
Root cause of a bug or failing test. Call it as soon as your first hypothesis fails; give it the symptom, the error output and the few files involved. Read-only.
Inputs: workdir (required), objective (required), diff, tier, +7 more.
codex_reviewcodex_reviewA
Adversarial review of a diff or a few files by Codex. Use after any non-trivial change. Returns a verdict and ranked findings with file:line, each checked against the files. Read-only.
Inputs: workdir (required), objective (required), diff, tier, +7 more.
councilcouncilA
The same question to Codex, Gemini and GPT-OSS in parallel, merged: each verdict, the findings two or more families agree on (the strongest signal), and where they disagree. Use for risky changes, security-sensitive code or a design you are unsure of. Read-only.
Inputs: workdir (required), objective (required), diff, mode, +10 more.
edge_casesedge_casesA
Inputs and states the code mishandles, phrased as test cases. Use after writing logic with branches, parsing, dates, money or retries. Read-only.
Inputs: workdir (required), objective (required), diff, tier, +7 more.
gauntlet_statusgauntlet_statusA
Which CLIs are installed and signed in, which model each tier uses, which models are paused after failures, and the limits in force. Call it after a failed call. Never shows credentials.
Inputs: no named parameters.
gemini_analyzegemini_analyzeB
How a broad or unfamiliar part of the code fits together. Use before editing code you have not read. Read-only.
Inputs: workdir (required), objective (required), diff, tier, +7 more.
gemini_reviewgemini_reviewB
Second review by Gemini, strongest on interface contracts, error paths, state and user-facing regressions. Read-only.
Inputs: workdir (required), objective (required), diff, tier, +7 more.
plan_critiqueplan_critiqueA
Critique of an implementation plan before any code is written: wrong assumptions, missing steps, simpler paths, production risks. Use on any task that touches more than one file. Read-only.
Inputs: plan (required), workdir (required), objective (required), diff, +8 more.
quick_checkquick_checkA
Fast second look at something small: a function, a regex, a query, a config value, a one-file change. Cheap and quick; use it often. Read-only.
Inputs: workdir (required), objective (required), diff, tier, +7 more.
security_auditsecurity_auditA
Exploitable issues: injection, authorization gaps, secrets, SSRF, path traversal, races, trust of client input. Use on auth, payments, user input, file paths, shell commands. Read-only.
Inputs: workdir (required), objective (required), diff, tier, +7 more.
Tool Schema Changelog
Recent tool additions, removals, and schema changes observed during successful MCP inspections.
Loading tool schema changelog…
TDQS
Scored across 10 tools
Most tools have clearly distinct purposes, such as status, quick check, security audit, edge cases, and plan critique. However, codex_review, gemini_review, and council overlap in the review space, and an agent might be unsure when to use a single-model review versus the multi-model council, despite the descriptions offering guidance.
All names use snake_case, which is consistent, but the underlying pattern is mixed: some are model_action (codex_review, gemini_analyze), some are noun_noun (security_audit, edge_cases, plan_critique), and others are adjective_noun (quick_check) or a single noun (council). This is readable but not a predictable verb_noun convention throughout.
The server provides 10 tools, which fits comfortably within the ideal range of 3–15 for a focused code-review and analysis assistant. Each tool appears to earn its place by covering a specific analysis angle or model combination.
The surface covers a broad range of code quality concerns: status, quick checks, reviews, diagnosis, architecture analysis, security, edge cases, plan critique, and multi-model council. Minor gaps exist, such as no explicit tool for performance analysis or dependency auditing, but the core lifecycle for reviewing and understanding code is well represented.
Related MCP Connectors
Multi-brand AI code review for Claude: up to 12 reviewers and a judge, no install.
Multi-LLM council: 25+ frontier models in parallel, consensus scoring, verdict-first code review.
Multi-model code review: a panel of models + detectors return a pass/fail verdict. Paid via x402.
Code review by AI models from different companies, usually two. Shows where they agree and disagree.
Related MCP Servers
- AlicenseAqualityDmaintenanceEnables OpenAI Codex to consult Claude Code for co-analysis, adversarial second opinions, and read-only file review over the Model Context Protocol, with Claude remaining advisory and unable to modify files.923 npmMIT
- AlicenseNot gradedqualityCmaintenanceUses Claude as an independent second-model reviewer inside Codex, performing reviews on disposable Git snapshots with structured findings.MIT

megalens-mcpofficial
AlicenseNot gradedqualityCmaintenanceEnables multi-model AI code review from inside supported coding tools, reporting where up to four models agree and disagree.36 npmMIT- AlicenseNot gradedqualityCmaintenanceEnables users to delegate coding tasks in a chosen project to the official Codex and Claude CLI agents from within an existing chat, then convene a council of configured models to gather independent opinions and review the resulting files and test runs. It also manages multi-provider API connections, model catalogs, and per-task token, time, and cost limits.MIT