ClawCoat v11.0.1 - Public Launch

[QuietFireAI]

The earned autonomy platform for AI agents. Self-hosted. Trust is earned by agents, granted by humans.

746 tests. 788 proof documents. 61,278 lines scanned. Apache 2.0.

Try the live demo: https://huggingface.co/spaces/QuietFireAI/ClawCoat

---

What Is ClawCoat?

A self-hosted, governance-first platform for autonomous AI agents. It acts as a governed MCP proxy: agents connect to ClawCoat, and every action they attempt is evaluated against trust levels, behavioral scoring, anomaly detection, and approval gates before execution. The agent is never modified. ClawCoat wraps it.

---

The Idea

Two things.

Agents earn their autonomy. Every agent starts at QUARANTINE - no tools, no external access, no assumptions. Five trust tiers: QUARANTINE -> PROBATION -> RESIDENT -> CITIZEN -> AGENT. Promotion is sequential and requires explicit human authorization. Demotion is instant and can skip levels. An agent cannot change its own trust tier.

Behavior has a score. Every agent carries a Manners Score - a live measurement (0.0-1.0) across five principles: Human Control, Transparency, Value Alignment, Privacy, and Security. The score moves with every action. Drop below 0.50: automatic demotion. Drop below 0.25: auto-suspend. No human delay required.

Everything else - the audit trail, the kill switches, the compliance frameworks, the 8-step governance pipeline - exists to support those two things.

---

The 8-Step Governance Pipeline

Every action, every time:

Step 1: Instance registered?       Reject if unknown
Step 2: Kill switch (suspended?)   Reject immediately
Step 3: Nonce replay protection    Reject if replayed
Step 4: Tool on blocklist?         Reject if blocked
Step 5: Classify action category   READ / WRITE / DELETE / EXTERNAL / FINANCIAL / SYSTEM
Step 6: Manners compliance score   Auto-quarantine if < 0.25 or 3+ violations / 24h
Step 7: Trust level permission     Allow / Gate / Block per matrix
Step 8: Anomaly detection          Flag behavioral deviations

The kill switch fires at Step 2 - before trust levels, before Manners, before everything except "does this agent exist?" One API call suspends any agent. Only a human can reinstate it.

---

What Is Shipped

Capability	Tests
Trust Level Governance - 5-tier earned trust, sequential promotion, instant demotion	54
Cryptographic Audit Trail - SHA-256 hash-chained, tamper-evident	11
150 RBAC Endpoints - 4-tier permissions, deny overrides allow	13
AES-256-GCM Encryption - at-rest, PBKDF2 key derivation	11
TOTP Multi-Factor Auth - RFC 6238, QR enrollment, backup codes	8
Behavioral Anomaly Detection - rate spikes, capability probes, enumeration	12
Human-in-the-Loop Gates - approval workflows with timeouts and escalation	9
Manners Compliance Engine - five behavioral principles, runtime scoring	7
Egress Firewall - domain whitelist, external call governance	5
Multi-Tenant Isolation - Redis key namespacing, litigation holds	8
OpenClaw Governance - governed MCP proxy, kill switch, Manners auto-demotion	55
Kill Switch - instant suspension, Redis-persisted, survives restarts	7
MCP Gateway - 13 tools, native Goose / Claude Desktop integration	live

746 tests passing. 1 skipped. 0 high-severity findings.

---

Compliance Frameworks

Framework	Coverage
SOC 2 Type I	51 controls documented, evidence mapped to source
HIPAA Security Rule	Administrative, Physical, Technical, Organizational
HITRUST CSF	12 domains, baseline controls, risk scoring
CJIS	Advanced auth, media protection, audit controls
GDPR	Data minimization, encryption, right to erasure
PCI DSS	Encryption, segmentation, access control, logging
ABA Model Rules	Rules 1.6, 1.7/1.10, 5.3, Formal Opinion 512

Every control references a source file and a passing test. Run proof_sheets/ to verify any claim.

---

Quick Start

git clone https://github.com/QuietFireAI/ClawCoat.git
cd ClawCoat
cp .env.example .env
# Edit .env: set MCP_API_KEY and JWT_SECRET_KEY
docker compose up --build -d
docker compose exec mcp_server alembic upgrade head
curl http://localhost:8000/health

Runs on a mini-PC, a Raspberry Pi, a homelab server, or a cloud VM. No subscription. No data harvesting. No terms that change.

---

Who This Is For

Law firms, healthcare, insurance, accounting - ClawCoat was built against the standards they operate under. HIPAA. SOC 2. HITRUST. CJIS. GDPR. PCI DSS. ABA Model Rules.

Small businesses get the same platform. No enterprise contract required.

Anyone running agents who wants to stay in control of their own data, on their own hardware.

---

This discussion was created from the release ClawCoat v11.0.1 - Public Launch.

[musaabhasan]

Earned autonomy is the right framing. The design detail I would watch closely is promotion evidence: agents should not advance because they completed many low-risk actions; they should advance because they completed representative actions for the tier they are trying to enter.

A strong scoring model could separate:

• reliability evidence: task completion, low retry loops, stable outputs
• security evidence: no policy bypass attempts, clean tool arguments, no credential exposure
• operator evidence: approvals granted, approvals denied, human correction frequency
• drift evidence: behavior change after model, prompt, tool, or memory updates

I would also make demotion faster than promotion. If an agent earns CITIZEN over weeks but then starts touching new tools, a new network domain, or a changed memory set, the tier should partially reset for that capability area rather than globally trusting the old score.

This keeps the platform explainable: "this agent is trusted for read-only ticket triage, still probationary for code changes, and quarantined for finance tools." That is much closer to how organizations actually manage risk.