getclaw 0.1.0

Clawrence — AI agent security framework. Any model, full audit trail, fail-closed by default.

ProClaw

Secure runtime for OpenClaw agents. Any model, full audit trail, fail-closed by default.

ProClaw is a security wrapper around OpenClaw that intercepts every tool call your agent makes and enforces policy before execution. It supports 10 model providers -- so if you're blocked by Anthropic or want cheaper inference, you can switch to DeepSeek, MiniMax, Together AI, Mistral, Gemini, Qwen, Llama, Ollama, OpenAI, or Azure OpenAI with one command.

Features

Security Stack (SteelClaw 8-Layer)

• 5-Level Security Mode — Safe → Strict → Guarded → Balanced → Autonomous
• Policy Engine — tool allowlists, filesystem permissions, network domain control, rate limits, time policies
• HITL Approval — human-in-the-loop for risky tool calls (CLI interactive, dashboard, mobile push)
• DLP Engine — Presidio + spaCy NER, regex patterns, credential detection, content redaction
• Audit Ledger — hash-chained tamper-evident event log with SQLite persistence
• Intent-Based Security — zero-trust middleware scoring agent intent (Launch 2)
• Shell Scanner — 7-threat classification for shell commands
• Sub-Agent Security — permission intersection (child ⊆ parent)

Agent Runtime

• Response Streaming — token-by-token output in CLI and SSE for dashboard/channels
• Typing Indicators — progress events at every stage (thinking, tool execution, completion)
• Tool Execution Progress — real-time SSE events: tool.started, tool.completed, tool.denied
• Context Management — sliding window + priority strategies with visible trimming notifications
• Tool Caching — TTL-based caching for repeated tool calls
• Retry with Escalation — exponential backoff with HITL escalation on repeated failures
• Credential Scrubbing — 8 regex patterns + exact match redaction on all tool output
• Prompt Injection Defense — input/output scanning for injection attempts

Sessions & Memory

• Conversation Sessions — SQLite-persisted multi-turn conversations across runs
• Semantic Memory — LanceDB embeddings with automatic fact extraction and retrieval
• Session Management — list, switch, delete, compact sessions from CLI

CLI

• Interactive Chat — proclaw chat REPL with streaming, sessions, history
• Model Management — proclaw models list, proclaw models recommend, proclaw models info
• Diagnostics — proclaw doctor checks Python, deps, config, providers, disk, DB
• Skill Management — proclaw skill install|list|info|remove|enable|disable|scan|verify
• Memory Management — proclaw memory list|search|add|forget|clear|stats
• Agent Control — proclaw kill for graceful stop during runs
• Channel Management — proclaw channels setup|start|stop|status|login

Messaging Channels

• Telegram — grammY adapter with inline buttons, streaming via message edits, silent messages, push debounce
• WhatsApp — Baileys adapter with read receipts, QR pairing, text-based approval
• Discord — discord.js adapter with button components, thread-per-run
• Channel Bridge — Node.js sidecar on port 9200, typing indicators, ack reactions, response pacing, welcome tour

Dashboard

• Real-Time Monitoring — live activity feed, approval queue, threat detection (wired to SQLite)
• Execution Replay — timeline view with play/pause, hash chain verification
• PWA Support — installable, service worker caching, push notifications, offline indicator
• Preview Badges — Launch 2 features clearly labeled with amber "Preview" banners

Model Support

• 10 providers: DeepSeek, Together, Anthropic, OpenAI, Google, Mistral, Ollama, MiniMax, Qwen, Azure
• Model aliases: opus, sonnet, deepseek, minimax, etc.
• Cost/quality comparison: proclaw models recommend

Install

pip install proclaw
proclaw setup

The setup wizard prompts you to pick a model provider and enter your API key. Then you're ready to run agents:

proclaw run agent.yaml -p "Read README.md and summarize it"

See the Quickstart Guide for the full walkthrough.

Why ProClaw?

OpenClaw gives agents powerful tools: file access, shell execution, browser control, API calls. Those tools are useful -- and dangerous. A misconfigured agent can exfiltrate data, delete files, or rack up unbounded API costs.

ProClaw sits between the agent and those tools. It reads a YAML manifest that declares what the agent is permitted to do, and denies everything else. The policy engine is fail-closed: if it can't reach a verdict, the answer is deny.

Every decision is recorded in an immutable, hash-chained audit log. You can verify the integrity of the chain at any time, and export it for compliance.

And because ProClaw is model-agnostic, you're never locked into a single provider. If Anthropic blocks your use case, switch to any of 10 supported providers in one command and keep running.

Documentation

Document	Description
Quickstart	3-minute setup guide
Model Comparison	Provider comparison table -- cost, speed, tool-calling quality
Agent Manifest Reference	Every field in agent.yaml explained

Architecture: SteelClaw 8-Layer Stack

L7  Operator Control Plane    --> dashboard/
L6  DLP Engine                --> dlp/
L5  Audit Ledger              --> audit/
L4  Network Enforcement       --> network/
L3  Zero-Trust Policy         --> policy_engine/, identity/
L2  Secure Runtime            --> interceptor/
L1  Model Isolation           --> sandbox/
L0  Deployment                --> deploy/

Every agent action passes through these layers. The Interceptor (L2) captures every OpenClaw tool call and routes it through the policy engine (L3) before allowing execution.

CLI Commands

proclaw setup                          # Interactive provider + API key wizard
proclaw config show                    # View current configuration
proclaw config set model <provider>    # Switch model provider
proclaw config set security-mode <lvl> # Set default security level
proclaw health                         # Check provider connectivity
proclaw doctor                         # Full diagnostic check
proclaw models list                    # Provider compatibility matrix
proclaw models recommend               # Best model recommendation
proclaw models info <provider>         # Detailed provider info
proclaw run agent.yaml -p "prompt"     # Run agent with manifest
proclaw run agent.yaml -p "prompt" --security-mode guarded --stream
proclaw chat agent.yaml -s my-session  # Interactive chat mode
proclaw sessions list                  # List conversation sessions
proclaw sessions delete <name>         # Delete a session
proclaw sessions compact <name>        # Remove old messages
proclaw kill                           # Stop running agent
proclaw skill list                     # List installed skills
proclaw skill install <path>           # Install from directory
proclaw skill scan <path>              # Security scan without install
proclaw memory list                    # List stored memories
proclaw memory search "query"          # Semantic search
proclaw memory add "fact"              # Remember a fact
proclaw channels setup                 # Configure messaging channels
proclaw channels start                 # Start channel bridge

Project Structure

ProClaw/
├── contracts/         Frozen integration interfaces (7 Protocols + 35 types)
├── interceptor/       OpenClaw tool call interception
├── sandbox/           Firecracker/gVisor/WASM isolation
├── network/           Envoy sidecar + DNS filtering
├── identity/          SPIFFE/SPIRE + Vault
├── policy_engine/     OPA + manifest + rate limits
├── dlp/               Presidio PII + spaCy NER
├── audit/             Hash-chained immutable ledger
├── dashboard/         Next.js operator UI
├── deploy/            Helm + Terraform + air-gap
├── enterprise/        SSO + RBAC + workspaces
├── lib/               Reference modules (zero_trust, streaming, etc.)
└── docs/              Documentation

Development

# Python environment
python -m venv .venv
source .venv/bin/activate  # or .venv\Scripts\activate on Windows
pip install -e ".[dev]"

# Run tests
pytest tests/ -v

# Dashboard
cd dashboard && npm install && npm run dev

# Full stack (Docker)
docker-compose up -d

License

Proprietary. All rights reserved.