Joint Chiefs

Multi-model AI code review. One command sends your code to OpenAI, Gemini, Grok, and Claude in parallel, runs a structured debate where they challenge each other's findings, and streams a single consensus summary back.

Website: jointchiefs.ai

$ jointchiefs review src/auth.swift --goal "security audit"

Why

Single-model code review has blind spots. GPT misses things Gemini catches; Gemini misses things Grok catches. Joint Chiefs runs them all, then has them argue.

The debate protocol is grounded in Multi-Agent Debate research (Liang et al., 2023), which shows that adversarial collaboration between LLMs produces more reliable output than any single model — including a single model reflecting on its own work.

Related MCP server: MCP Code Crosscheck

Surfaces

Joint Chiefs ships as three binaries, one engine:

A fourth binary, jointchiefs-keygetter, is the single binary that reads and writes the local API-key file. The CLI and MCP server call it via Process.

Requirements

• Apple Silicon Mac (M-series). Intel Macs are not supported.

• macOS 15 (Sequoia) or later.

• Xcode 16+ with the macOS 15 SDK.

• API keys for at least one supported provider.

Install

From source

git clone https://github.com/djfunboy/joint-chiefs.git
cd joint-chiefs/JointChiefs
swift build -c release
cp .build/release/jointchiefs .build/release/jointchiefs-mcp .build/release/jointchiefs-keygetter /opt/homebrew/bin/

API keys

Two paths:

a) Environment variables (CI-friendly):

export OPENAI_API_KEY="sk-..."
export GEMINI_API_KEY="..."
export GROK_API_KEY="..."
export ANTHROPIC_API_KEY="sk-ant-..."   # also acts as the moderator

b) Local credential file (end-user default, via the setup app):

Run jointchiefs-setup. It walks through disclosure, key entry (with live Test buttons), strategy config, install location, and outputs the MCP config snippet for your AI client. Keys are saved to ~/Library/Application Support/Joint Chiefs/credentials.json (mode 0600), so the CLI and headless agents read them with no prompt. Upgrading from v0.5.6 or earlier? The setup app migrates any keys from the old macOS Keychain automatically on first launch.

You only need one key to get started. More keys = more diverse debate.

Verify:

jointchiefs models

Usage

Review a file:

jointchiefs review src/auth.swift

With a directive to focus the panel:

jointchiefs review src/auth.swift --goal "look for race conditions in token refresh"

Pipe a diff in from stdin:

git diff main | jointchiefs review --stdin --goal "pre-commit check"

Quiet mode (suppress streaming, print only the final consensus):

jointchiefs review src/auth.swift --quiet

JSON output:

jointchiefs review src/auth.swift --format json

MCP integration

The MCP server (jointchiefs-mcp) works with any MCP-aware client. The setup app's MCP Config tab emits a ready-to-paste mcpServers JSON snippet keyed at the installed binary path. No keys live in the snippet — Joint Chiefs resolves them from the local credential file at tool-call time.

Minimal snippet shape:

{
  "mcpServers": {
    "joint-chiefs": {
      "command": "/opt/homebrew/bin/jointchiefs-mcp"
    }
  }
}

How it works

                  ┌──────────────────────┐
                  │  DebateOrchestrator  │
                  └──────────┬───────────┘
                             │
       ┌──────────┬──────────┼──────────┬──────────┐
       ▼          ▼          ▼          ▼          ▼
   ┌────────┐┌──────────┐┌────────┐┌────────┐┌─────────┐
   │ OpenAI ││Anthropic ││ Gemini ││  Grok  ││  local  │   parallel review
   └────┬───┘└────┬─────┘└───┬────┘└───┬────┘└────┬────┘   (Ollama / LM Studio
        └────────┴──────────┼──────────┴──────────┘         / any OpenAI-compat)
                            │  anonymized findings
                            ▼
                  ┌────────────────────┐
                  │      Moderator     │   synthesizes round, writes brief
                  │  (default: Claude) │
                  └─────────┬──────────┘
                            │
                            ▼
             Next round, or — if positions converged —
             the moderator writes the final consensus.

You pick which providers participate. The moderator is configurable too (default: Claude); the same provider can serve as both a spoke and the moderator if you want, or you can split the roles.

Up to 5 debate rounds with adaptive early break when positions converge. Findings are anonymized before the final synthesis to reduce bias toward any single provider. Four consensus modes (moderatorDecides, strictMajority, bestOfAll, votingThreshold) with per-provider weighting.

Full details: docs/ARCHITECTURE.md.

Configuration

CLI flags:

Privacy

• API keys live in a permission-locked local file (~/Library/Application Support/Joint Chiefs/credentials.json, mode 0600), encrypted at rest by FileVault and read only via the jointchiefs-keygetter binary. Env vars exist as a CI fallback.

• No telemetry. No analytics. The only network traffic is to the LLM APIs you've configured.

• The MCP server is stdio-only — nothing binds a port.

• Code sent for review is stored only in local transcript files. Delete them whenever.

Development

cd JointChiefs
swift test          # current test count changes over time
swift build -c release

The project layout:

JointChiefs/
├── Package.swift
├── Sources/
│   ├── JointChiefsCore/        # Models, providers, orchestrator, APIKeyResolver
│   ├── JointChiefsCLI/         # jointchiefs executable
│   ├── JointChiefsMCP/         # jointchiefs-mcp stdio server
│   ├── JointChiefsSetup/       # jointchiefs-setup SwiftUI installer
│   └── JointChiefsKeygetter/   # jointchiefs-keygetter — sole credential-file accessor
└── Tests/JointChiefsCoreTests/

Contributing

PRs welcome — especially first-class providers for Mistral and DeepSeek. Both are reachable today via the OpenAI-compatible path (point OPENAI_COMPATIBLE_BASE_URL at https://api.mistral.ai/v1 or https://api.deepseek.com/v1); native providers would add curated model lists and provider-specific handling. This is a solo-maintained project, so response times will vary. No SLA, no promises about backwards compatibility, and breaking changes can land on main between releases.

If you want to add a provider, conform to the ReviewProvider protocol in Sources/JointChiefsCore/Services/Providers/ReviewProvider.swift and use SSE streaming end-to-end — non-streaming LLM calls are not accepted.

License

MIT — see LICENSE.