Platform features
OpenPawz ships with more capabilities than any other AI agent platform — open source or commercial. No other tool combines multi-agent orchestration, 11 simultaneous chat channels, multi-chain DeFi trading, semantic memory with hybrid retrieval, voice, a visual node editor, and a three-tier extension marketplace — all running locally on your desktop with zero cloud dependency. This is the complete feature inventory.Agents
Custom personalities
Each agent has a name, bio, system prompt, emoji, and avatar. Three personality axes — tone (casual → formal), initiative (reactive → proactive), and detail (brief → thorough) — let you fine-tune behavior without prompt engineering.
Agent templates
Pre-built archetypes — General, Research, Creative, Technical, Custom — so you’re productive in seconds. Templates set personality, system prompt, and tool access automatically.
75+ built-in tools
Agents call tools autonomously — send emails, search the web, execute shell commands, browse pages, trade crypto, talk to other agents. All built in, not plugins.
Soul files
Persistent identity documents for each agent. Soul files store personality notes, custom instructions, and identity data that persist across sessions. Your agent evolves over time.
Per-agent workspaces
Every agent gets an isolated filesystem workspace with file listing, browsing, and cleanup. Agents can’t accidentally interfere with each other’s files.
Agent boundaries
Define what each agent should and shouldn’t do. Boundary rules constrain behavior without complex prompt engineering — just plain language rules.
Per-agent model override
Each agent can use a different AI model, overriding the global default. Run your research agent on Claude, your coder on GPT-4o, and your assistant on a local Ollama model — simultaneously.
Auto-approve mode
Trusted agents can skip human-in-the-loop approval prompts entirely. Toggle per agent — full autonomy for agents you trust, strict oversight for new ones.
11 simultaneous channels
No other AI agent platform connects to this many platforms at once. One app, one set of agents, eleven live bridges — running simultaneously. A user messages you on Telegram, another on Discord, a third on Slack — the same agent (or different agents) respond instantly on each platform.| Channel | Type | Key capabilities |
|---|---|---|
| Telegram | Bot API | DM policy, user allowlists, pairing approval |
| Discord | Bot | Mention-response mode, DM policy, user approval |
| Slack | Socket Mode | App + bot tokens, mention-response, DM policy |
| Matrix | Client | Homeserver auth, room response, DM policy |
| IRC | TLS Client | Multi-channel join, channel response toggle |
| Web bridge | Evolution API, Docker management, QR pairing, group response | |
| Twitch | Chat | OAuth bot, multi-channel, mention-require toggle |
| Mattermost | Bot | Server + token auth, mention-response, DM policy |
| Nextcloud Talk | Bot | Username/password auth, group response toggle |
| Nostr | Relay | NIP-04 encrypted DMs, Ed25519 signing, multi-relay WebSocket |
| Webchat | Built-in | TLS WebSocket, cookie auth, inline HTML/JS/CSS, session management |
Uniform access control
Every channel has the same interface: start/stop, config get/set, status, approve/deny/remove users. Consistent behavior across all 11 platforms.
Per-channel agent routing
Assign a specific agent to each channel. Your support bot handles Discord, your research agent handles Slack, your personal assistant handles Telegram — all from one app.
Channel routing rules
Rule-based routing engine with first-match logic. Route messages to specific agents based on channel, user, keywords, or custom conditions.
Smart message splitting
Long responses automatically split at platform-appropriate boundaries. Discord’s 2000-char limit, Telegram’s 4096-char limit — handled transparently.
AI providers
Bring your own model. OpenPawz supports 10 providers — local and cloud — and you can switch between them per agent, per task, or per conversation.| Provider | Type | Models |
|---|---|---|
| Ollama | Local | Llama, Mistral, Gemma, Phi, Qwen, and hundreds more |
| OpenAI | Cloud | GPT-4o, o1, o3, GPT-4.1 |
| Anthropic | Cloud | Claude 4, Sonnet, Haiku |
| Cloud | Gemini 2.x | |
| DeepSeek | Cloud | DeepSeek V3, R1 |
| Grok | Cloud | Grok 3 |
| Mistral | Cloud | Mistral Large, Codestral |
| OpenRouter | Cloud | Multi-model gateway |
| Moonshot | Cloud | Kimi |
| Custom | Any | Any OpenAI-compatible endpoint |
Model routing
Boss model vs worker model assignment. OpenPawz automatically routes simple queries to cheaper models and complex tasks to powerful ones — saving you money without sacrificing quality.
SSE streaming
All providers stream responses via Server-Sent Events with real-time deltas. You see tokens appear as the model thinks — no waiting for the full response.
Thinking & reasoning blocks
Full support for Anthropic thinking blocks and Google thought handling. See the model’s reasoning process in real time.
Daily budget enforcement
Set a hard USD spending cap per day. When the budget is reached, API calls are blocked. Never get surprised by a bill again. No other platform has this.
Tools — 75+ built-in
Not plugins. Not add-ons. Built directly into the platform. Agents can use any combination of these tools autonomously.- Core tools
- Agent tools
- Trading tools
- Voice tools
- Community tools
| Tool | What it does |
|---|---|
| Shell execution | Run arbitrary commands with risk classification |
| File operations | Read, write, append, delete, list files in agent workspaces |
| Web fetch | HTTP fetch with configurable domain allowlists/blocklists |
| Browser automation | Headless Chrome — navigate, screenshot, extract, click, fill forms |
| Full IMAP/SMTP: read, send, search, folder management, AI draft generation | |
| GitHub | Repo operations, issues, PRs, code search |
| Memory | Store, search, list, delete semantic memories |
| Tasks | Create, update, move tasks on the Kanban board |
Extensibility — three tiers + marketplace
OpenPawz has the most sophisticated extension system of any agent platform. Three tiers let you start simple and graduate to full apps — plus a community marketplace.- Skills (Tier 1)
- Integrations (Tier 2)
- Extensions (Tier 3)
- MCP
- PawzHub marketplace
Drop a markdown file. That’s it. A
SKILL.md file injects expert knowledge into an agent’s system prompt. No code, no config, no API — just markdown. Hot-reloaded from ~/.paw/skills/ with no restart needed.Agent-installable skills
No other platform has this. Agents can conversationally search PawzHub, discover new capabilities, and install skills for themselves — at runtime. Your AI agent literally improves itself.
Skill Wizard
AI-assisted TOML manifest generation. Paste an API’s documentation and the wizard generates a complete integration manifest automatically.
TOML hot reload
Drop a new
.toml file into ~/.paw/skills/ and it’s instantly available. No restart, no rebuild, no deployment. The filesystem is your package manager.Per-agent skill scoping
Skills, integrations, and extensions can be enabled for specific agents only. Your research agent gets web tools, your support agent gets email tools — no overlap.
Multi-agent orchestration
Boss/worker delegation
A boss agent plans a project and delegates tasks to specialist workers. Workers run in parallel, report results, and the boss synthesizes the output. Full lifecycle: planning → running → paused → completed → failed.
Agent squads
Unique to OpenPawz. Persistent groups of agents with coordinator and member roles. Create a “Research Squad” or “Content Team” and reuse them across projects. Squads coordinate automatically.
Agent specialties
Assign workers a specialty — coder, researcher, designer, communicator, security, general. Each worker can have its own model and system prompt within a project.
Inter-agent messaging
Agents send direct messages to specific agents or broadcast to all. Full delegation, progress, result, and error message streams between boss and workers.
Task management
Full Kanban board with columns: inbox, assigned, in progress, review, blocked, done. Tasks have priorities (low → urgent), multi-agent assignment with lead/collaborator roles, and per-task model overrides.
Automations
Cron scheduling with 60-second heartbeat, event-driven triggers, persistent execution mode, and a full cron run log with audit trail. Schedule agents to run research at 2 AM, post updates at 9 AM, clean up at midnight.
Memory — the most advanced retrieval system in any agent platform
Most AI tools forget everything between conversations. OpenPawz has the most sophisticated memory system available — four retrieval strategies combined.Hybrid retrieval
No other platform does this. BM25 full-text search + vector similarity (cosine) combined with weighted merging. You get the precision of keyword search and the flexibility of semantic search — together.
MMR re-ranking
Maximal Marginal Relevance with Jaccard-based diversity ensures recalled memories are both relevant AND diverse. No more getting five copies of the same memory.
Temporal decay
Exponential decay with a 30-day half-life. Recent memories rank higher than old ones. Your agent prioritizes what’s current without losing long-term knowledge.
Auto-capture & auto-recall
Key facts are automatically extracted from every conversation and stored with vector embeddings. When context is relevant, memories are injected before the agent responds — zero manual effort.
Memory Palace
A dedicated interface to browse, search, edit, and delete memories. Filter by agent, topic, date, category, or relevance score. Full transparency into what your AI knows.
Importance scoring
Each memory receives an importance score (0–1). Critical facts rank higher than casual observations. The system learns what matters.
Memory backfill
Retroactively generate vector embeddings for existing memories. Upgrade your memory system without starting over.
Shared & isolated namespaces
All agents share memory by default — knowledge captured by one is available to all. Isolate per agent or per workspace when you need boundaries.
Session management & token savings
Automatic session compaction
When conversations approach the context limit, OpenPawz summarizes older messages, preserves key facts in memory, archives the full session, and continues seamlessly. Agents run indefinitely long conversations — and you pay for 40–60% fewer tokens.
Thinking level control
Unique to OpenPawz. Adjust reasoning depth per session — none, low, medium, high. Need a quick answer? Use low. Complex analysis? Crank it up. You control the compute.
Configurable context windows
Set the context window size in tokens (default 32,000). Larger windows for complex tasks, smaller ones for cost savings. Per-session control.
Max concurrent runs
Limit simultaneous agent runs across chat, cron, and tasks. Chat always gets priority. Prevent runaway costs from parallel automations.
DeFi trading — multi-chain, built-in
No other AI agent platform has built-in, multi-chain cryptocurrency trading. OpenPawz agents can trade on Coinbase, Solana (via Jupiter/Pump.fun), and Ethereum (via Uniswap) — autonomously, with configurable safety guardrails.Coinbase exchange
Full exchange trading via Coinbase API: limit/market orders, account balances, candle data, trade history.
Solana DEX
Jupiter aggregator for best-price swaps, Pump.fun integration for memecoins, Ed25519 wallet management, SOL/SPL token transfers, portfolio tracking.
Ethereum DEX
Uniswap V2/V3 swaps, full ABI/RLP encoding, HD wallet (BIP-39/44), ERC-20 operations, token discovery, liquidity analysis.
Position monitoring
Track open positions with stop-loss and take-profit targets. Real-time price feeds across chains. Aggregate portfolio view.
Trading policy engine
Unique to OpenPawz. Configurable auto-approve rules with max trade size (USD), daily loss limits, allowed pairs whitelist, and transfer caps. Agents trade autonomously within guardrails you set.
Trade audit trail
Every trade logged: type, side, amount, USD value, reason, agent, session, timestamp. Full accountability for autonomous trading decisions.
Voice — text-to-speech & speech-to-text
Three TTS providers
Choose between Google TTS, OpenAI TTS, or ElevenLabs. Configurable voice selection, speed, and language per agent.
Speech-to-text
Audio transcription via Whisper-compatible engines. Speak to your agents and they understand.
ElevenLabs integration
Advanced voice controls: model selection, stability sliders, similarity boost. Professional-grade voice output.
Auto-speak
Toggle automatic voicing of all new agent responses. Your agents become voice assistants — read responses aloud as they arrive.
Security — zero-trust, local-first
OpenPawz was designed from the ground up with a zero-trust, local-first security model. This isn’t an afterthought bolted on — it’s in the architecture.100% local — zero cloud
All data stays on your hardware. No cloud sync, no telemetry, no analytics, no phone-home. The only outbound traffic is to AI providers you explicitly configure.
OS keychain credentials
API keys stored in macOS Keychain, Windows Credential Manager, or Linux Secret Service. Never in plain text. Never in a config file.
AES-256-GCM encryption
Database fields encrypted with AES-256-GCM using 12-byte random nonces. Auto-migration from legacy encryption. Your data is encrypted at rest.
Command risk classification
Every tool and shell command rated 1–5 for risk. 30+ danger pattern detections: privilege escalation, destructive deletion, remote code execution, network exfiltration.
Human-in-the-loop
Modal-based approval for all side-effect tools. Color-coded risk levels. Nothing dangerous happens without your explicit confirmation.
Timed session override
Unique to OpenPawz. “Allow All” mode for trusted sessions (30m/1h/2h) with auto-expire. Privilege escalation is still blocked — even in override mode.
Container sandbox
Run agent shell commands in Docker containers — cap_drop ALL, memory/CPU limits, configurable network isolation. Agents can’t touch your host system.
Per-agent policies
Granular control over which tools, channels, and risk levels each agent can access.
Prompt injection defense
Dual implementation (TypeScript + Rust), 30+ detection patterns, 4 severity levels. Incoming messages are analyzed before they reach agents.
Filesystem sandboxing
Tauri scope enforcement with 20+ sensitive path blocks —
~/.ssh, ~/.aws, ~/.gnupg, and more. Optional global read-only mode blocks all agent writes.Outbound domain control
Configurable domain allowlists and blocklists with wildcard subdomain matching. Test URL button to verify rules. Outbound requests inspected for exfiltration patterns.
Credential audit trail
Unique to OpenPawz. Every credential access logged: action, tool, allowed/denied, timestamp. Full security audit dashboard with JSON/CSV export.
Skill vetting pipeline
Unique to OpenPawz. Community skills go through a safety modal, known-safe list check, npm registry risk intelligence scan, and post-install sandbox verification.
CSP enforcement
Content Security Policy:
default-src 'self', script-src 'self', object-src 'none', frame-ancestors 'none'. XSS and injection prevention at the platform level.Visual node editor
Unique to OpenPawz. A visual, drag-and-drop workflow editor for building agent pipelines without writing code. Connect nodes — triggers, agents, tools, conditions, outputs — to create complex automation flows visually. → Deep dive: Projects guideFoundry — skill creation workshop
The Foundry is a dedicated workspace for creating and testing skills, integrations, and extensions. AI-assisted manifest generation, live preview, and one-click publishing to PawzHub. → Deep dive: Foundry guideToday View — your AI dashboard
A daily activity dashboard that shows everything your agents have been doing. Activity feed, quick actions, inbox, and status across all agents, channels, and automations — at a glance. → Deep dive: Dashboard guideEmail — full IMAP/SMTP client
Not a toy integration — a complete email client. IMAP/SMTP via himalaya bridge: read, send, search, folder management, AI-powered draft generation. Multiple accounts, folder operations, and a dedicated Mail view in the app. → Deep dive: Email guideRemote access
Tailscale integration
Unique to OpenPawz. Expose your agents remotely via Tailscale mesh VPN. Serve and Funnel support for secure remote access — no port forwarding, no cloud relay. Connect from anywhere.
Webhook server
Built-in HTTP webhook endpoint with auth token, rate limiting, and configurable agent routing. Trigger agents from GitHub Actions, Zapier, IFTTT, or any HTTP client.
Developer experience
20 slash commands
Built-in slash commands with real-time autocomplete. Quick access to common operations without leaving the chat.
Modular workspace
Toggle grid for showing/hiding UI views. Community extensions add to this grid. Customize the interface for your workflow.
15 settings tabs
Agent defaults, models, sessions, skills, MCP, browser, engine, environment, logs, advanced, voice, webhook, Tailscale, and more. Everything is configurable.
Persistent structured logging
Daily log rotation with 7-day pruning and an in-app log viewer with filtering. Debug anything without digging through files.
Architecture & quality
Native Tauri v2
Rust backend + TypeScript frontend. Not Electron — fast, lean, native. 158 IPC commands covering every engine operation.
Zero-framework frontend
Vanilla TypeScript with atomic design (atoms → molecules → organisms). No React, no Vue, no Angular. Minimal bundle, maximum performance.
530 automated tests
164 Rust + 366 TypeScript tests covering all critical paths. 3-job CI pipeline (Rust + TypeScript + Security Audit) on every push.
Zero known CVEs
cargo audit + npm audit enforced in CI. Zero Clippy warnings. Dependabot for automated dependency updates.21 SQLite tables
Full relational schema for agents, sessions, memories, tasks, trades, skills, channels, and more. All local, all fast.
Retry with circuit breakers
Exponential backoff (1s → 30s, 3 retries),
Retry-After header support, 5-failure circuit breaker with 60-second cooldown. Resilient by default.What exists like this? Nothing.
| Capability | OpenPawz | OpenClaw | AutoGPT | CrewAI | LangChain |
|---|---|---|---|---|---|
| Desktop native app | Tauri/Rust | Electron | No — CLI/web | No — Python lib | No — Python lib |
| Chat channels | 11 simultaneous | 1–2 | 0 | 0 | 0 |
| Multi-agent orchestration | Boss/worker + squads | Basic | Single agent | Yes | Yes (framework) |
| Semantic memory | Hybrid BM25+vector+MMR+decay | Basic | Basic | None built-in | Framework only |
| Built-in tools | 75+ | ~20 | ~15 | Code-defined | Framework only |
| DeFi trading | 3 chains built-in | None | None | None | None |
| Voice (TTS/STT) | 3 providers | None | None | None | None |
| Extension marketplace | 3 tiers + PawzHub | None | Plugin store | None | LangSmith |
| Self-installing skills | Agents install their own | No | No | No | No |
| Visual node editor | Built-in | No | No | No | LangGraph Studio |
| Session compaction | Automatic, 40–60% savings | No | No | No | No |
| Daily budget caps | Built-in USD cap | No | No | No | No |
| Security audit trail | Full + export | Basic | None | None | None |
| Webhooks + Tailscale | Built-in | None | None | None | None |
| Credential security | OS keychain + AES-256 | Config files | Config files | Env vars | Env vars |
| Open source | MIT | Proprietary | MIT | MIT | MIT |