@cofounder/ci

CI/CD guardrails for AI applications. Scan your codebase for security, compliance, cost, and safety issues in AI/LLM code -- with zero runtime dependencies.

Features

• No Hardcoded Keys - Detect API keys, secrets, and credentials (Anthropic, OpenAI, AWS, Groq, xAI)
• No PII in Prompts - Find personally identifiable information in prompt templates and test fixtures
• No Injection Vulnerabilities - Detect prompt injection patterns (string concatenation, f-strings, unsanitized input)
• Approved Models - Enforce an allowlist of approved LLM models, warn on deprecated models
• Cost Estimation - Estimate monthly costs from LLM calls found in code
• Safe Defaults - Check for unsafe configs (temperature > 2, missing max_tokens, missing system prompts)

Installation

npm install @cofounder/ci --save-dev

CLI Usage

# Scan current directory
npx cofounder-ci scan

# Scan specific path with specific rules
npx cofounder-ci scan ./src --rules no-hardcoded-keys,no-injection-vuln --format json

# Fail only on critical issues
npx cofounder-ci scan . --fail-on critical

# Output SARIF for GitHub Security tab
npx cofounder-ci scan . --format sarif > results.sarif

# Validate configuration
npx cofounder-ci validate .cofounder.yml

# Scan and post to GitHub PR
npx cofounder-ci check . --format github-pr

GitHub Action Usage

name: AI Guardrails
on:
  pull_request:
    branches: [main]

jobs:
  scan:
    runs-on: ubuntu-latest
    steps:
      - uses: actions/checkout@v4

      - name: CoFounder CI Scan
        uses: waymaker-ai/cofounder/packages/ci@v1
        with:
          scan-path: '.'
          rules: 'all'
          fail-on: 'high'
          config: '.cofounder.yml'
          comment-on-pr: 'true'
          github-token: ${{ secrets.GITHUB_TOKEN }}
          approved-models: 'claude-sonnet-4-20250514,gpt-4o,gemini-2.0-flash'
          budget-limit: '500'

Action Inputs

| Input | Description | Default | |-------|-------------|---------| | scan-path | Path to scan | . | | rules | Comma-separated rules to enable | all | | fail-on | Severity to fail on (critical, high, medium, low, none) | high | | config | Path to .cofounder.yml config | .cofounder.yml | | comment-on-pr | Whether to comment on PRs with results | true | | github-token | GitHub token for PR comments | ${{ github.token }} | | approved-models | Comma-separated list of approved models | (all common models) | | budget-limit | Monthly budget limit in USD | (none) |

Action Outputs

| Output | Description | |--------|-------------| | findings | Number of findings | | critical | Number of critical findings | | passed | Whether the scan passed |

Configuration (.cofounder.yml)

rules:
  no-hardcoded-keys:
    enabled: true
    severity: critical
  no-pii-in-prompts:
    enabled: true
    severity: high
  approved-models:
    enabled: true
    severity: medium

scan:
  exclude:
    - "dist"
    - "node_modules"
    - "*.test.ts"

models:
  approved:
    - claude-sonnet-4-20250514
    - gpt-4o
    - gemini-2.0-flash
  blocked:
    - text-davinci-003

budget:
  monthly: 500
  perCall: 0.05

ignore:
  - "*.md"
  - "docs/"

.cofounderignore

Create a .cofounderignore file in your project root (gitignore-style syntax):

# Ignore test fixtures with intentional secrets
test/fixtures/
*.test.ts

# Ignore generated files
dist/
build/

Output Formats

• console - ANSI-colored terminal output (default)
• json - Machine-readable JSON
• sarif - SARIF 2.1.0 for GitHub Security tab
• markdown - Markdown report
• github-pr - GitHub PR comment with tables and expandable details

Programmatic Usage

import { scan, formatReport } from '@cofounder/ci';

const result = scan({
  scanPath: './src',
  rules: 'all',
  failOn: 'high',
  format: 'json',
  ignorePatterns: [],
  commentOnPr: false,
});

console.log(`Found ${result.findings.length} issues`);
console.log(`Passed: ${result.passed}`);

const report = formatReport(result, 'json');
console.log(report);

Rules

| Rule | Severity | Description | |------|----------|-------------| | no-hardcoded-keys | critical | Detect hardcoded API keys and secrets | | no-pii-in-prompts | high | Detect PII in prompt templates | | no-injection-vuln | critical | Detect prompt injection vulnerabilities | | approved-models | medium | Enforce approved model list | | cost-estimation | medium | Estimate LLM usage costs | | safe-defaults | medium | Check for unsafe LLM configurations |

License

MIT