NVIDIA Inception · 9 Backends · Policy-Driven Security

Autonomous AI Agents
Enterprise-Ready

Deploy, secure, and scale autonomous AI agents across Telegram, Discord, WhatsApp, and email. 9 inference backends. Declarative security policies. Full audit trails. Your infrastructure, your rules.

pureclaw
9
Backends
4
Channels
22
Observers
990+
Tests
6
Security Modules

One Command. Full Control.

The pureclaw CLI gives you direct control over your agent from the terminal. Start channels, manage observers, validate security policies, check status.

pureclaw start

Launch the agent with your chosen engine. All channels, observers, and security modules boot in seconds.

pureclaw start --engine hybrid

pureclaw status

Live view of running observers, active channels, engine health, and policy state. One glance, full picture.

pureclaw status --verbose

pureclaw policy

Validate and hot-reload security policies without restarting. Lint YAML, test rules, preview enforcement.

pureclaw policy validate --strict
Install
$ pip install pureclaw $ pureclaw init # Or clone the source $ git clone https://github.com/puretensor/PureClaw.git $ cd PureClaw && pip install -e .

Production-Grade Agent Infrastructure

Multi-channel ingress, pluggable inference backends, autonomous monitoring, and a heuristic dispatcher for sub-second data retrieval.

Multi-Engine Inference

9 backends across 3 tiers: local GPU (vLLM, Ollama), cloud API (Anthropic, Bedrock, Gemini), and CLI agents (Claude Code, Codex, Gemini CLI). Hybrid router for automatic failover. Switch with one environment variable.

ENGINE_BACKEND=hybrid python3 nexus.py

Multi-Channel Ingress

Telegram, Discord, WhatsApp, and Email. Each channel maintains isolated sessions with persistent context. Voice input via Whisper transcription. Human-in-the-loop approval for outbound actions.

4 channels · per-user session isolation

Observer System

22 autonomous cron-scheduled monitors running in a ThreadPoolExecutor. Email digest, morning brief, cyber threat feeds, git security audit, Darwin rail data consumer, infrastructure health, and daily reporting.

Internal scheduler · no external crontab

Heuristic Dispatcher

10 pre-LLM data APIs for sub-second structured responses. Weather, markets, crypto, forex, gold, trains, Darwin, commute, infrastructure, and system status. Regex-matched commands bypass inference entirely.

Pattern match → direct API → data card

Policy-Driven Security. Auditable by Design.

6 security modules. Declarative YAML policies with hot-reload. Full audit trails. 104 security-specific tests.

Policy Engine

Declarative YAML security policies. Define allowed commands, filesystem paths, network destinations, and inference constraints. Hot-reload without restart. Draft policy recommendations generated from denied actions.

security/policy.py · hot-reload

Audit Trail

Structured logging of every agent action, tool invocation, and policy decision. Denial aggregation with frequency analysis. Exportable compliance records for regulatory review.

security/audit.py · structured logs

Credential Redaction

Automatic detection and redaction of API keys, tokens, passwords, and secrets in agent output and logs. Pattern-based scanning before any content reaches channels or storage.

security/redact.py · pre-output scan

SSRF Protection

Network-level controls preventing server-side request forgery. Allowlist and denylist for outbound connections. Internal network ranges blocked by default. All egress logged.

security/network.py · egress control

Filesystem ACLs

Path-level access control for agent file operations. Define readable, writable, and forbidden paths in policy YAML. Prevents lateral movement across filesystem boundaries.

security/filesystem.py · path ACLs

Inference Guards

Controls over model selection, token limits, and prompt injection defences. Restrict which models the agent can invoke. Rate limiting and cost controls for API-backed engines.

security/inference.py · model controls
security/policy.yaml Example Policy
security: filesystem: allowed_read: ["/opt/app", "/var/data"] allowed_write: ["/opt/app/output"] denied: ["/etc", "/root", "~/.ssh"] network: allow_egress: ["api.anthropic.com", "api.openai.com"] deny_private: true # Block RFC1918 ranges inference: max_tokens: 8192 allowed_models: ["claude-*", "gpt-4*"] redaction: enabled: true # Scan all output patterns: ["api_key", "token", "password"]

9 Backends. 3 Tiers. Zero Lock-in.

Local GPU, cloud API, or CLI agents. Switch with one environment variable or a tap in Telegram. Hybrid router for automatic failover.

Local GPU
Cloud API
CLI Agents
vllm Local GPU
Default inference backend
OpenAI-compatible endpoint. Nemotron, Qwen, LLaMA, Mistral, DeepSeek. Continuous batching, tensor parallelism, speculative decoding.
19 tools · Free
ollama Local GPU
Local inference
Any GGUF model. Lightweight deployment for edge and development. Your hardware, your data, no network required.
19 tools · Free
anthropic Cloud API
Anthropic API
Direct Anthropic API access. Claude Sonnet, Opus, Haiku. Native tool use with structured responses.
19 tools · API key
bedrock Cloud API
AWS Bedrock
Claude models via AWS Bedrock. IAM-based auth, VPC endpoints, CloudTrail logging. Enterprise AWS integration.
19 tools · AWS IAM
gemini Cloud API
Gemini API
Google Gemini models via API. Gemini Pro, Flash, and Ultra. Native function calling support.
19 tools · API key
claude_code CLI Agent
Claude Code CLI
Full agentic tool use. File editing, bash execution, search. The complete Claude Code experience, orchestrated by PureClaw.
Delegated tools · Subscription
codex_cli CLI Agent
Codex CLI
OpenAI Codex CLI agent. Full agentic tool use for code generation and system operations.
Delegated tools · Subscription
gemini_cli CLI Agent
Gemini CLI
Google Gemini CLI agent with full tool use. Agentic coding and system tasks.
Delegated tools · Subscription
hybrid Router
Automatic failover
Routes between API and CLI backends based on availability, latency, and cost. Automatic fallback when primary engine is unavailable.
Auto-routing · Mixed

Production-Ready in Minutes

Install, init, start. Runs as a systemd service or Kubernetes pod with full observability.

bash
$ pip install pureclaw $ pureclaw init # Configure channels, credentials, security policy $ pureclaw policy validate # Choose backend: vllm, ollama, anthropic, bedrock, gemini, # claude_code, codex_cli, gemini_cli, hybrid $ pureclaw start --engine hybrid

Built for Enterprise Procurement

Security controls, audit trails, and policy enforcement designed to meet the requirements of regulated environments.

ISO 27001

Policy engine, audit logging, access controls, and incident recording aligned with ISMS requirements.

Cyber Essentials

Boundary firewalls (SSRF protection), access control (filesystem ACLs), secure configuration (YAML policies), malware protection (credential redaction).

G-Cloud Ready

Self-hosted deployment model. Full audit trail export. Policy-as-code for reproducible security posture. Open source for transparency.

Zero Telemetry

No data leaves your infrastructure. No analytics, tracking, or usage reporting. MIT licensed. Fully auditable source code.

Early Access

Stay in the loop.

Get notified when new engines, features, and integrations drop. No noise, just signal.

You're on the list.

No spam. Unsubscribe anytime.