rigour

Rigour

Your AI agent just tried to commit an AWS secret. Rigour blocked it in <100ms.

Agent Transaction Firewall (v6): treat the agent as an untrusted proposer. Rigour decides—deterministically—what it may write, run, call, and ship. No AI judge on the allow/deny path.

Try it now (zero config)

npx rigour-scan

Works on any repo. No init, no config, no setup. Instant results in your terminal:

  HARDCODED SECRET DETECTED
  AWS_SECRET_ACCESS_KEY found in src/config.ts:23

  + 22 more violations across 847 files (2.1s)

  Score        ████░░░░░░░░░░░░░░░░  34/100
  AI Health    ███░░░░░░░░░░░░░░░░░░  28/100

  Gates:  ✅ file-size  ❌ security  ❌ ast  ✅ deps

  Brain: learned 12 patterns · trend: improving ↑

Add to your AI IDE (30 seconds)

{ "mcpServers": { "rigour": { "command": "npx", "args": ["-y", "@rigour-labs/mcp@latest"] } } }

IDE / Agent	MCP Tools	Live Dashboard	Real-Time Feed
Claude Desktop	✅	✅ MCP App	✅ Logging
VS Code Copilot	✅	✅ MCP App	✅ Logging
ChatGPT	✅	✅ MCP App	✅ Logging
Goose	✅	✅ MCP App	✅ Logging
Claude Code	✅	—	✅ Logging
Cursor	✅	—	✅ Logging
Cline	✅	—	✅ Logging
Windsurf	✅	—	✅ Logging
Codex	✅	—	✅ Logging

Then install hooks so writes are checked in real time:

npx @rigour-labs/cli hooks init --tool cursor   # or claude, cline, windsurf
# or via MCP: rigour_hooks_init

Does the firewall run automatically?

Short answer: quality gates + DLP + mediated rigour_run paths run when MCP/hooks are installed. It does not yet sit in front of every third-party MCP tool (GitHub/Slack/etc.)—that gateway is designed but not the default proxy.

Surface	Automatic once installed?	What you get
MCP server (@rigour-labs/mcp)	Yes for Rigour tools the agent calls	rigour_check, Fix Packets, memory DLP, agent register, Studio events
rigour_run / rigour_run_supervised	Yes when those tools are used	Typed argv allowlist (no free-form shell: true), fail-closed human arbitration (timeout = deny), one-time Studio token
IDE hooks (Cursor/Claude/Cline/Windsurf)	Yes after hooks init	Per-write checks (secrets, imports, size, protected paths). If agents are registered, set RIGOUR_AGENT_ID so scope binds to the writer (fail-closed; no union-allow)
rigour_agent_register	Yes when called	Rejects **/* / sensitive globs unless operator scopes or RIGOUR_ALLOW_AGENT_SCOPE_AUTHORITY=1
CLI firewall	On demand / CI	firewall adversarial, firewall transact, firewall admit
Third-party MCP proxy	Not yet	Capability broker + McpGateway interfaces exist; Rigour is not yet a MITM for all MCP servers

Agent proposes action
        ↓
Hooks / MCP mediation (when on the path)
        ↓
Deterministic deny/allow + rule id
        ↓
Studio evidence  ·  CI attestation (admit)

Agent Transaction Firewall

npx @rigour-labs/cli firewall status
npx @rigour-labs/cli firewall adversarial   # deterministic corpus — unexpected allows fail CI
npx @rigour-labs/cli firewall transact --agent <id> --scope 'packages/foo/**'
npx @rigour-labs/cli firewall admit         # CI: valid attestation + PASS + bound git tree
npx @rigour-labs/cli studio                 # Firewall tab: decisions, attestation, mediation health

Guarantees (on mediated paths): fail-closed arbitration · typed commands · per-agent scope · signed attestation bound to commit/tree · adversarial replay as regression fuel—not an AI red-team product.

See ADR 001.

Live governance dashboard (MCP App)

In supported editors, a real-time dashboard appears automatically as your agent works:

┌─ Rigour Governance ──────────────────────────┐
│  Score: 94/100  ✅ PASS                      │
│                                               │
│  14:32:01  rigour_check → FAIL (34/100)       │
│  14:32:03  fix_packet → 8 fixes               │
│  14:32:15  rigour_check → 71/100 (+37)        │
│  14:32:22  rigour_check → ✅ PASS 94/100      │
│                                               │
│  Brain: 47 patterns · trend: improving ↑      │
└───────────────────────────────────────────────┘

No extra commands. The dashboard appears when the agent calls Rigour tools. Watch your agent self-heal in real time. Open Firewall in Studio for allow/deny decisions and mediation status (partial until the MCP gateway is fully wired).

What it catches

Category	Gates
Security	Hardcoded secrets (29+ patterns), SQL injection, XSS, CSRF, prototype pollution, Shannon entropy
Structural	File size, cyclomatic complexity, method count, parameter count, nesting depth, TODO/FIXME
AI Drift	Hallucinated imports, phantom APIs, context drift, retry loop detection
Governance	Agent team isolation, checkpoint supervision, memory DLP
Firewall	Out-of-scope writes, undeclared MCP tools, disallowed shell, fail-closed timeouts

AST-based. Not heuristics. TypeScript, JavaScript, Python, Go, Ruby, C#, Java, Kotlin, Rust.

How it works

Agent writes code → Hooks / gates fire → FAIL? → Fix Packet (JSON)
                                           ↓
                                    Agent reads exact instructions
                                           ↓
                                    Agent fixes → PASS ✓

Mediated run (rigour_run) → typed allowlist → human arbitration (fail-closed)
                                           ↓
                                    execute or deny + evidence

Voluntary rigour_check is a quality workflow, not the security boundary. Hard guarantees require installed hooks/MCP mediation and (for CI) firewall admit.

The Brain — learns your codebase

Every scan reinforces patterns. Patterns decay when absent. At strength: 0.9, they promote to hard rules. Your project's own immune system — trained locally, zero telemetry.

First week:  catches 12 violations
First month: catches 8 violations  ← learning your patterns
Third month: catches 3 violations  ← your agents have adapted

How it's different

	Rigour	ESLint	“AI security agents”
Runs locally, zero telemetry	✅	✅	often ❌
Learns YOUR codebase (Brain)	✅	❌	❌
Agent self-healing (Fix Packets)	✅	❌	❌
Deterministic execution firewall	✅	❌	usually LLM judge
Works offline (GGUF sidecar)	✅	✅	❌
AI-native drift detection	✅	❌	❌
MCP-native	✅	❌	varies

Used in production

• 19,000+ total installs across CLI and MCP
• Organically forked by Alibaba iFlow
• OWASP project — listed
• Cursor MCP directory — listed

Quick reference

npx rigour-scan                              # zero-config scan
npx @rigour-labs/cli init                    # add gates to your project
npx @rigour-labs/cli hooks init --tool cursor
npx @rigour-labs/cli check                   # run gates
npx @rigour-labs/cli check --deep            # + local AI analysis
npx @rigour-labs/cli check --deep --provider claude -k sk-ant-xxx  # cloud AI
npx @rigour-labs/cli studio                  # monitoring + Firewall tab
npx @rigour-labs/cli firewall adversarial
npx @rigour-labs/cli firewall admit

Architecture

Package	Purpose
@rigour-labs/core	Gate engine, AST, Fix Packets, Brain, firewall kernel
@rigour-labs/cli	init, check, scan, run, studio, firewall
@rigour-labs/mcp	MCP server — governance tools for agent integration
rigour-scan	Zero-config shortcut: npx rigour-scan

Stack: TypeScript strict, web-tree-sitter, Zod, Vitest.

---

Full docs | Technical Spec | Philosophy | Firewall ADR

MIT © Rigour Labs — Built by Ashutosh

If Rigour caught something real in your codebase — tell us.