The operating system for autonomous AI agents. Multi-channel, multi-agent, safety-first. Every action follows a strict six-stage pipeline — the LLM reasons, the system executes.
Every action — without exception — passes through six validated stages. The LLM generates plans. The system executes them. They never cross.
Sanitize input. Rate limit. Detect prompt injection across 15+ patterns.
LLM generates strict JSON plan. No execution, no side effects, reasoning only.
Schema check. Safety policy. Blockchain simulation. Risk scoring.
Tools, APIs, browser, blockchain. Sandboxed and scoped. No raw LLM access.
Confirm results. Self-heal failures. Diagnose → fix → retry → escalate.
Full Clawtrace audit trail. Auto-redacted secrets. JSONL structured output.
Production-grade components that work together. Extend with skills, MCP servers, or custom tools.
Per-user mode toggle. Low-risk actions auto-execute, critical actions always confirm. Safety pipeline runs in both modes.
Puppeteer-based control. Navigate, click, type, extract, screenshot. Full headless browser in the agent pipeline.
Agent registry with capability routing. Delegate tasks, coordinate work, enforce depth limits across sub-agents.
Persistent semantic memory. TF-IDF embeddings with cosine similarity search. Scoped by session, user, or global.
Transfers, balance checks, SPL tokens, transaction simulation. On-chain agent registry with PDA-derived identity.
Connect to any Model Context Protocol server. Discover and invoke external tools without writing code.
Visual workflow engine. Triggers, conditions, parallel branches, error handling. No external orchestrator needed.
AES-256-GCM encryption at rest. Ed25519 signing. Keys zeroed after use. Never logged, never leaked.
Intelligent failure recovery. Diagnose root cause, apply fix strategy, retry intelligently, escalate only when needed.
Eleven channels. One brain. All sharing the same tools, memory, and safety pipeline. Install only what you need.
Telegram via Telegraf · Discord via discord.js · Slack via Bolt · WhatsApp via Baileys · Email via nodemailer + IMAP · SMS via Twilio · WebChat via WebSocket · Webhooks via HTTP · LINE via Messaging API · Reddit via OAuth2 · Matrix via Client-Server API
PAW agents execute JavaScript, TypeScript, Python, Rust, and any language you throw at them via the api, js, and system execution modes.
But PAW is the only agent framework that also ships with a Solana-native smart contract language built in — Purp SCL v1.1.
Write, compile, and deploy Solana programs without leaving the chat.
program TokenVault { } account VaultState { owner: pubkey balance: u64 is_locked: bool } instruction Deposit { accounts: #[mut] vault_state #[signer] depositor args: amount: u64 body: require(amount > 0, InsufficientFunds) vault_state.balance += amount emit(DepositMade, { depositor: depositor.key, amount: amount }) } event DepositMade { depositor: pubkey amount: u64 } error VaultErrors { InsufficientFunds = "Not enough funds" Unauthorized = "Only the owner" }
Safety-first architecture where the LLM plans and the system executes.
No exceptions, no shortcuts, every action traceable.
The validation pipeline runs on every action, in every mode. You can make it less intrusive with autonomous mode, but you can't turn it off.
The LLM produces a JSON plan. The system validates it, then executes. The LLM never touches tools directly — eliminating an entire class of prompt injection attacks.
Autonomous doesn't mean unsafe. Low and medium risk auto-execute. High risk is configurable. Critical actions always require confirmation.
Clawtrace logs every phase of every action with automatic secret redaction. Reconstruct exactly what happened, what the LLM reasoned, and why.
When an action fails, PAW diagnoses the failure type, determines if it's recoverable, applies a fix strategy, and only escalates when it genuinely can't recover.
Parse .purp files, validate, compile to Anchor Rust, generate TypeScript SDKs and IDL — all within the agent pipeline. No SCL integration like it.
Vector memory with TF-IDF scoring, scoped by session, user, or global, persisted to disk. Agents remember what matters across conversations.
Skill files, MCP protocol, custom tools, DAG workflows, multi-agent orchestration. PAW grows with your needs without code changes.
A built-in web dashboard served by the Gateway — monitor agents, chat in real-time, switch modes, and inspect every action as it happens.
Talk to your agents over WebSocket. Messages appear instantly with user/agent/system bubbles.
Toggle between supervised, autonomous, and locked modes on the fly — no restart needed.
Action count, uptime percentage, and pipeline status update live via WebSocket push.
Every action logged with timestamps and color-coded severity — info, success, warning, error.
Full theme toggle with persistent preference. Purple brand palette across both modes.
Served automatically by the PAW Gateway. Open your browser and the dashboard is ready.
Clone, configure, run. One AI API key and one channel token to start.
# Clone the repository $ git clone https://github.com/DosukaSOL/paw-agents.git $ cd paw-agents # Install dependencies $ npm install # Configure your API keys and channel tokens $ cp .env.example .env # Build and start $ npm run build $ npm start
Requires Node.js 20+, at least one AI API key (OpenAI or Anthropic), and at least one channel token.
Six providers with automatic failover. Configure your preferred model — use one or all.
Automatic failover between all providers · ModelProvider interface for extensibility