triage-ai

Collaborative AI triage for code, plans and architecture — Claude, Gemini and Codex independently explore your codebase, then merge findings with consensus scoring into one prioritized report. Use it for security audits, bug hunts, plan reviews, second opinions on AI-generated code, and more. Available as a Claude Code skill or MCP server for any AI editor.

npm install -g triage-ai
triage-ai "find security vulnerabilities"

How It Works

1. Each AI agent explores your codebase independently — reading files, following imports, searching for patterns. They're not limited to what the scanner finds; they run in read-only mode and can investigate anything in your repository.
2. Findings are merged — identical issues from 2+ models become consensus findings (high confidence, low false positives). Severity disagreements are flagged as conflicts for you to decide.
3. One prioritized report — S0 (blockers) through S3 (style), with patches in unified diff format.

You need at least one AI CLI installed. Three catch what one misses.

What It Does

• Security audits — SQL injection, XSS, command injection, auth bypass, hardcoded secrets
• Bug detection — logic errors, race conditions, null references, missing error handling
• Second-opinion on AI plans — before implementing what Claude/Gemini/Copilot proposed, get independent multi-model review
• Architecture analysis — agents explore the full codebase to understand context, not just individual files
• Consensus scoring — when multiple models independently flag the same issue, confidence is high
• Conflict detection — when models disagree on severity, you get both perspectives
• Auto-patching — models propose unified diffs, applied safely on a new git branch (--dry-run to preview)
• AI memory — --remember saves findings to CLAUDE.md / GEMINI.md / AGENTS.md so every AI tool in your project knows what triage found

Safety Model

Each AI CLI runs in non-interactive pipe mode — they can read and explore your codebase but cannot modify it:

| CLI | Mode | What Happens | |-----|------|-------------| | Claude Code | -p (print mode) | Reads files, searches, analyzes — no interactive writes | | Gemini CLI | -p (print mode) | Reads files, searches, analyzes — no interactive writes | | OpenAI Codex | --sandbox read-only | Full auto-approve in a read-only sandbox — cannot write |

Secrets (API keys, passwords, private keys, AWS credentials, GitHub tokens, npm tokens, Slack tokens, Stripe keys, Anthropic/OpenAI keys) are redacted from any context sent to the models. Sensitive files (.env, credentials, binaries) are excluded automatically.

Install

npm install -g triage-ai
triage-ai setup              # detects installed CLIs, offers to install missing ones
triage-ai ready              # smoke test — verifies each model can respond

Requires Node.js 18+ and at least one AI CLI:

| Model | Install | Auth | |-------|---------|------| | Claude Code | npm install -g @anthropic-ai/claude-code | claude (interactive login) | | Gemini CLI | npm install -g @google/gemini-cli | gemini (interactive login) | | OpenAI Codex | npm install -g @openai/codex | codex (interactive login) |

Usage

cd your-project

# Full 3-model review
triage-ai "find bugs and security issues"

# Single model, quick check
triage-ai --models claude "quick security scan"

# Review only uncommitted changes
triage-ai --diff-only "check my changes for bugs"

# Save report + remember findings
triage-ai --remember --out report.md "full security audit"

# Preview patches without applying
triage-ai --dry-run "fix the SQL injection"

# Restrict models to pre-gathered context only (faster, no filesystem exploration)
triage-ai --context-only "review this code for issues"

Second-opinion on AI plans

A standout use case: get a multi-model review of plans proposed by an AI before you implement them.

When Claude, Gemini, Copilot or any AI coding assistant proposes a plan — a refactor, a new feature, an architecture change — run triage to get independent second opinions before committing to it:

# After Claude proposes changes in plan mode
triage-ai "Claude proposed the following plan — review it for correctness, \
  security risks, edge cases and anything it might have missed: \
  [paste or describe the plan]"

# Review AI-generated code that's been staged
triage-ai --diff-only "an AI generated these changes — review for bugs and security"

# Validate a migration plan
triage-ai "Review this database migration plan for data loss risks, \
  missing rollback steps and performance issues"

Each model independently evaluates the plan against your actual codebase — exploring files, checking assumptions, and flagging issues the proposing AI may have overlooked. Consensus findings (flagged by 2+ models) are especially worth paying attention to.

Example Output

TTY mode (interactive terminal)

┌ triage-ai v1.3.0
│
├ Intake
│  ✓ Scanned repository          42 files, 3 modified
│  ✓ Built context package        186 KB across 28 files
│
├ Triage Team
│  ✓ Claude                       found at /usr/local/bin/claude v2.1.50
│  ✓ Gemini                       found at /usr/local/bin/gemini v0.30.0
│  ✗ Codex                        not installed (skipping)
│
├ Assessment
│  ✓ Claude                       16 findings (47.3s)
│  ✓ Gemini                       14 findings (38.2s)
│
├ Diagnosis
│  ✓ Clustered findings           22 unique issues from 2 models
│  ✓ Consensus detected           4 issues confirmed by 2+ models
│
├ Report
│  ✓ Generated report             3 blockers, 5 high, 8 medium, 6 low
│
└ Done in 52.3s — 22 findings, 4 consensus

Non-TTY mode (CI / AI orchestrators)

When piped or run by an AI orchestrator (e.g. Claude Code), triage-ai outputs machine-parseable markers:

=== triage-ai v1.3.0 ===

[phase:1/6] intake — Intake
[intake] Scanning repository…
[intake] Scanning repository ✓ (42 files)
[intake] Built context package ✓ (186 KB across 42 files)

[phase:2/6] team — Triage Team
[team] Claude ✓ (found at /usr/local/bin/claude v2.1.50)
[team] Gemini ✓ (found at /usr/local/bin/gemini v0.30.0)
[team] Claude ✓, Gemini ✓

[phase:3/6] assess — Assessment
[assess] Claude…
[assess] Gemini…
[assess] Claude… 15s
[assess] Gemini… 15s
[assess] Gemini ✓ (14 findings (38.2s))
[assess] Claude ✓ (16 findings (47.3s))

[phase:4/6] diag — Diagnosis
[diag] Clustered findings ✓ (22 unique issues from 2 models)
[diag] Consensus detected ✓ (4 issues confirmed by 2+ models)

[phase:5/6] report — Report
[report] Generated report ✓ (3 blockers, 5 high, 8 medium, 6 low)
[report] 3 S0, 5 S1, 8 S2, 6 S3

=== TRIAGE COMPLETE ===
Time: 52.3s | Findings: 22 | Consensus: 4
Severity: 3 blockers, 5 high, 8 medium, 6 low

Model Results:
  ✓ claude   16 findings in 47.3s
  ✓ gemini   14 findings in 38.2s
======================

=== REPORT START ===
# Triage Report
...
=== REPORT END ===

CLI Reference

triage-ai [PROMPT] [OPTIONS]
triage-ai setup                  # detect CLIs, install missing, show auth hints
triage-ai ready [models]         # smoke test all or specific models

| Option | Default | Description | |--------|---------|-------------| | --models | claude,gemini,codex | Which models to use (comma-separated) | | --diff-only | off | Only analyze git diff | | --max-files | 200 | Max files in initial context (agents explore beyond this) | | --context-only | off | Restrict models to pre-gathered context (faster, no exploration) | | --format | md | Output format: md or json | | --out | stdout | Write report to file | | --apply | off | Apply patches (creates git branch first) | | --dry-run | off | Preview patches without applying | | --timeout | 300 | Per-model timeout in seconds | | --nice | 10 | Nice level for subprocess priority | | --remember | off | Save findings to CLAUDE.md, GEMINI.md, AGENTS.md | | --forget | — | Remove triage findings from memory files | | --verbose | off | Detailed progress output | | --mcp | — | Start as MCP server |

Severity Levels

| Level | Name | Examples | |-------|------|---------| | S0 | Blocker | SQL injection, RCE, auth bypass, data loss | | S1 | High | Logic bugs, race conditions, XSS | | S2 | Medium | N+1 queries, missing validation | | S3 | Low | Naming conventions, dead code |

AI Editor Integration

triage-ai works in any AI editor — as a Claude Code skill (auto-discovered, richest experience) or as an MCP server (universal, works everywhere).

Claude Code (Skill — recommended)

triage-ai ships with a SKILL.md that Claude Code discovers automatically after install. Claude will know when and how to run triage without any configuration.

npm install -g triage-ai
# That's it — ask Claude to "run a triage" or "find security issues"

You can also use the /triage slash command:

cp examples/claude-code-skill.md ~/.claude/commands/triage.md
# Then: /triage "find security vulnerabilities"

Claude Code (MCP)

Alternatively, add triage as an MCP server in ~/.claude/settings.json:

{
  "mcpServers": {
    "triage": {
      "command": "triage-ai",
      "args": ["--mcp"]
    }
  }
}

Gemini CLI

Add to ~/.gemini/settings.json:

{
  "mcpServers": {
    "triage": {
      "command": "triage-ai",
      "args": ["--mcp"]
    }
  }
}

For project-level context, copy the example memory file:

cp examples/gemini-md-example.md your-project/GEMINI.md

OpenAI Codex

Add to your Codex MCP configuration, or run directly:

codex "run triage-ai to find bugs"

For project-level context, copy the example memory file:

cp examples/agents-md-example.md your-project/AGENTS.md

Cursor / Windsurf / Cline / VS Code (Copilot) / Zed / Continue

Add to your editor's MCP configuration:

{
  "mcpServers": {
    "triage": {
      "command": "triage-ai",
      "args": ["--mcp"]
    }
  }
}

Configuration

Model overrides

export TRIAGE_CLAUDE_CMD="claude"              # override CLI command
export TRIAGE_GEMINI_CMD="gemini"
export TRIAGE_CODEX_CMD="codex"
export TRIAGE_GEMINI_MODEL="gemini-2.5-pro"    # override Gemini model

Other environment variables

export TRIAGE_HEARTBEAT_MS=15000               # non-TTY heartbeat interval (default 15s)

Disclaimer

triage-ai is a wrapper that orchestrates third-party AI CLI tools. By using it, you acknowledge:

• Your code is sent to external AI services (Anthropic, Google, OpenAI) via their respective CLI tools. Each service's terms of service and privacy policy apply.
• AI analysis is not a substitute for professional security audits. Findings may contain false positives or miss real vulnerabilities. Always verify AI-generated findings and patches before applying them.
• No warranty. This tool is provided as-is under the MIT license. The authors are not responsible for any damages, data loss, or security incidents arising from its use.
• Patches are best-effort. Always review AI-suggested patches before applying. Use --dry-run first.

Development

git clone https://github.com/wyman101/triage-ai.git
cd triage-ai && npm install && npm run build
npm test                     # runs vitest (18 tests)

License

MIT