Security scanner for AI-generated apps. Catches the bugs Cursor, Lovable, Windsurf, v0, and Claude routinely ship: hardcoded LLM keys, prompt-injection sinks, leaked Server Actions, hallucinated imports, missing auth on streaming endpoints, and the other "looks fine to a linter" issues that traditional tools miss.

# One-shot, no install
npx ubon@latest check

# Or install globally
npm install -g ubon
ubon check

ubon check                       # fast static scan, exit 1 on errors
ubon scan --interactive          # walk through findings one by one
ubon check --json                # deterministic JSON for agents/CI
ubon rules list --json           # machine-readable rule catalog
ubon check --sarif out.sarif     # SARIF 2.1.0 for GitHub code scanning
ubon mcp                         # serve as an MCP tool to your AI assistant
ubon doctor                      # check environment and optional deps

Modern AI coding assistants are great at producing code that runs. They are routinely careless about code that's safe to deploy:

• Hardcoded LLM API keys in client bundles
• Server Actions with no auth check
• Streaming routes with no rate limit
• MCP server configs with literal secrets
• import.meta.env.PUBLIC_* reading server-only values
• 'use client' files importing from actions/
• Edge runtime routes calling Node-only APIs
• Hallucinated imports that pass the type checker because the package never gets installed

Ubon's job is to catch those, fast, with high confidence and file:line context — and to expose them to the agent itself via JSON / NDJSON / MCP so the AI can fix what it broke.

v3.2.0 is an additive release for agentic development workflows: installable guardrails, richer machine-readable output, and a validation harness that proves Ubon catches planted AI-era bugs before a release ships.

• Agent harness installer: ubon agent install --all --write can generate Cursor, Claude Code, Codex, pre-commit, GitHub Actions, and .gitignore harness files from one dry-run-first workflow.
• Expanded Cursor hooks: templates now cover file edits, shell commands, MCP calls, prompt submission, stop gates, and pre-compaction context.
• Agent-specific rules (CC009–CC011): catches unknown Cursor hook events, broad agent autonomy, and dangerous reusable commands / skills.
• Agent-ready CLI: ubon changed, ubon verify, ubon review, ubon rules list --json, and presets for agent, ci, release, and local workflows.
• MCP upgrade: tools for changed-file scans, baseSha, verification, status, rule catalog access, and fix planning.
• Repair context: JSON / MCP output can include source context so agents have enough local evidence to patch findings.
• Validation harness: fixture benchmarks, CLI/MCP contract tests, deterministic fix/rescan checks, dogfood, and package dry-run verification are wired into npm run verify:release.
• Release discipline: npm run dogfood scans Ubon itself and must pass with 0 unsuppressed critical findings before publish.

For the original v3 breaking changes (Node 20+, removed Python / Rails / Vue profiles), see MIGRATION-v3.md.

Use them together. ESLint covers code style; npm audit covers CVEs in your dependency tree; Ubon covers the gap that AI assistants regularly leave behind.

ubon agent install --cursor --write   # writes Cursor hooks + rules

Then point Cursor at the MCP server:

// ~/.cursor/mcp.json
{
  "mcpServers": {
    "ubon": { "command": "npx", "args": ["-y", "ubon@latest", "mcp"] }
  }
}

Full Cursor + Lovable + comparison details in docs/INTEGRATIONS.md.

The demo fixture in examples/ai-harness-demo contains the kinds of issues AI agents often leave behind: an LLM key in source, server-side fetch to a user-controlled URL, a misspelled Cursor hook, and a reusable agent command that pipes network output into a shell.

ubon check -d examples/ai-harness-demo --preset local

The expected rule IDs are checked by the test suite so the demo stays honest.

ubon init                          # writes ubon.config.json
ubon check --update-baseline       # accept current findings as baseline
ubon check --baseline .ubon-baseline.json --focus-new --fail-on error

// ubon.config.json
{
  "profile": "next",
  "minConfidence": 0.85,
  "failOn": "error",
  "disabledRules": ["VIBE003"],
  "exclude": ["legacy/**"]
}

For the JS variant (executes user code), pass --allow-config-js or set UBON_ALLOW_CONFIG_JS=1.

• docs/CLI.md — every command and flag
• docs/START-HERE.md — two-minute setup and daily commands
• docs/AGENT-HARNESS.md — Cursor / Claude Code / Codex / MCP / hooks
• docs/AGENT-SEMANTICS.md — how Ubon models agent events, rules, and gates
• docs/PROGRAMMATIC.md — Node API, JSON/NDJSON, and MCP contracts
• docs/RULES.md — full rule glossary
• docs/CONFIG.md — config file schema
• docs/INTEGRATIONS.md — Cursor / Lovable / comparison
• docs/MCP.md — Model Context Protocol server
• docs/ADVANCED.md — profiles, suppressions, baselines, output schemas, release policy
• docs/RELEASE.md — release verification and publish checklist
• docs/VALIDATION.md — dogfood, fixtures, contracts, and repair loops
• MIGRATION-v3.md — upgrading from v2.x
• CHANGELOG.md — release history

• Node.js 20 or newer (v3 dropped Node 16/18)
• Git (for --git-changed-since and the git-history scanner)
• Optional: @modelcontextprotocol/sdk for ubon mcp — installed automatically as an optionalDependency of ubon. If your install flags skipped it, see docs/MCP.md.

Run ubon doctor to verify.

I'm Luisfer Romero Calero. I built Ubon because the gap between "AI shipped this" and "this is safe to deploy" keeps widening. The tool's name comes from the lotus (อุบล) in Thai — clarity in the middle of vibe-coded chaos.

If Ubon helps you ship safer apps, the highest praise is to wire it into your CI and your AI assistant — and tell me what it caught.

MIT — see LICENSE.