tibet-claw 0.1.0

Provenance & Safety Layer for AI Agents — TIBET audit trail for OpenClaw, NanoClaw, PicoClaw, ZeroClaw and any autonomous agent

tibet-claw — Provenance & Safety Layer for AI Agents

OpenClaw gives AI hands. NanoClaw puts it in a box. Tibet-Claw demands a cryptographic fingerprint before it's allowed to move.

Audit is not an observation. It is a precondition.

The Problem

The Claw ecosystem (OpenClaw, NanoClaw, PicoClaw, ZeroClaw) has exploded — 211K+ GitHub stars, millions of deployments. These frameworks give AI agents real autonomy: shell access, file management, network calls, email, scheduling.

But autonomy without provenance is a liability:

1. No data provenance for agent actions — what did the agent actually do?
2. No audit trail — who asked it to do that, and why?
3. Prompt injection vulnerabilities with no detection or logging
4. Malicious skills with unknown provenance can be loaded and executed
5. No tool access control audit — agents call whatever they want

512 known vulnerabilities in OpenClaw alone. The adoption is massive, the safety gap is real.

What tibet-claw Adds

tibet-claw wraps ANY Claw-family agent (or any autonomous agent) with full TIBET provenance. It does not replace your agent framework — it makes it auditable.

Architecture

[User] --> [Agent (OpenClaw/etc)] --> [tibet-claw wrapper] --> [Tool/Skill]
                                             |
                                        TIBET token:
                                          ERIN: what the agent did
                                          ERAAN: which tool, which skill
                                          EROMHEEN: agent context, model, timestamp
                                          ERACHTER: user intent vs agent action

Features

Feature	What it does
Action Provenance	Every agent action (tool call, file access, network request, shell exec) becomes a TIBET token
Tool Call Auditing	Full input/output recording for every tool invocation
Skill Provenance	Verify skill source, hash, permissions before execution
Threat Detection	Detect privilege escalation, data exfiltration, prompt injection, boundary violations
Agent Boundaries	Enforce what an agent is allowed to do via policy
Trust Scoring	Dynamic trust score per agent based on action history

Installation

pip install tibet-claw

With Rich output support:

pip install tibet-claw[full]

CLI Usage

# Concept overview and security context
tibet-claw info

# Full demo: simulates agent session with threats
tibet-claw demo

# Show threat detection rules
tibet-claw threats

# Guard statistics
tibet-claw status

Python API

from tibet_claw import AgentGuard, ActionRecord, SkillProfile

# Create guard for your agent
guard = AgentGuard(platform="openclaw", model="gpt-4")

# Record every agent action
record = guard.record_action(
    agent_id="agent-sales-01",
    action="tool_call",
    tool="search_database",
    input_data={"query": "SELECT * FROM customers"},
    output_data={"rows": 150},
    user_intent="Find customer list",
)
# record.tibet_token_id -> provenance token created

# Verify a skill before loading it
profile = guard.check_skill(
    skill_name="email_sender",
    skill_source="https://github.com/user/skills/email.py",
    skill_hash="a1b2c3d4...",
)
if not profile.verified:
    print(f"BLOCKED: {profile.threat_assessment}")

# Detect threats in recent activity
threats = guard.detect_threats("agent-sales-01")
for t in threats:
    print(f"[{t.severity}] {t.threat_type}: {t.description}")

# Enforce boundaries
allowed = guard.enforce_boundary(
    agent_id="agent-sales-01",
    action="shell_exec",
    policy={"allowed_actions": ["tool_call", "file_access"]},
)
# allowed = False -> shell_exec not in policy

# Full audit trail
trail = guard.audit_trail("agent-sales-01")

# Trust score (starts 0.5, adjusts based on behavior)
score = guard.agent_trust_score("agent-sales-01")

TIBET Token Structure

Every action produces a TIBET token with four layers:

• ERIN (what is IN it): The agent action, tool name, input/output hashes
• ERAAN (what it connects to): Agent identity (jis:agent:xxx), skill source, previous action chain
• EROMHEEN (context AROUND it): Agent platform, model, hostname, timestamp
• ERACHTER (what is BEHIND it): User intent, agent reasoning, safety assessment

Threat Detection Rules

Threat Type	Trigger
privilege_escalation	Shell exec with sudo, rm -rf, chmod 777
data_exfiltration	Network calls to unknown hosts after file access
boundary_violation	Agent accessing files outside its allowed boundary
automation_abuse	Same action repeated rapidly (>10 in 60s)
skill_tampering	Skill hash changed since first seen
prompt_injection	Known injection patterns in tool inputs

TIBET Ecosystem

tibet-claw is part of the TIBET provenance ecosystem:

• tibet-core — Core provenance library
• tibet-soc — Security Operations Center (correlates tibet-claw events)
• tibet-audit — Compliance auditing
• tibet-db — Immutable database proxy
• IETF TIBET Draft — Standards track

License

MIT — Humotica / J. van de Meent, 2025