openclaw.net

<div align="center">
  <img src="src/OpenClaw.Gateway/wwwroot/image.png" alt="OpenClaw.NET Logo" width="180" />
</div>

# OpenClaw.NET

[![License: MIT](https://img.shields.io/badge/License-MIT-blue.svg)](https://opensource.org/licenses/MIT)
![NativeAOT-friendly](https://img.shields.io/badge/NativeAOT-friendly-blue)
![Plugin compatibility](https://img.shields.io/badge/plugin%20compatibility-practical-green)
![Tools](https://img.shields.io/badge/native%20tools-48-green)
![Channels](https://img.shields.io/badge/channels-9-green)

> **Disclaimer**: This project is not affiliated with, endorsed by, or associated with [OpenClaw](https://github.com/openclaw/openclaw). It is an independent .NET implementation inspired by their work.

Self-hosted **AI agent runtime and gateway for .NET** with 48 native tools, 9 channel adapters, multi-agent routing, review-first self-evolving features, built-in OpenAI/Claude/Gemini provider support, built-in tool presets, NativeAOT support, and practical OpenClaw ecosystem compatibility.

## Why This Project Exists

Most agent stacks assume Python- or Node-first runtimes. That works until you want to keep the rest of your system in .NET, publish lean self-contained binaries, or reuse existing tools and plugins without rebuilding your runtime around another language stack.

OpenClaw.NET takes a different path:

- **NativeAOT-friendly runtime and gateway** for .NET agent workloads
- **Practical reuse of existing OpenClaw TS/JS plugins and `SKILL.md` packages** — install directly with `openclaw plugins install`
- A real **tool execution layer** with approval hooks, timeout handling, usage tracking, and optional sandbox routing
- **48 native tools** covering file ops, sessions, memory, web search, messaging, home automation, databases, email, calendar, and more
- **9 channel adapters** (Telegram, SMS, WhatsApp, Teams, Slack, Discord, Signal, email, webhooks) with DM policy, allowlists, and signature validation
- **Native LLM providers out of the box** for **OpenAI**, **Claude / Anthropic**, and **Gemini**, plus Azure OpenAI, Ollama, and OpenAI-compatible endpoints
- **Review-first self-evolving workflows** that can propose profile updates, automation drafts, and skill drafts from repeated successful sessions
- A foundation for **production-oriented agent infrastructure in .NET**

If this repo is useful to you, please star it.

## Key Features

### Agent Runtime
- Multi-step execution with tool calling, retries, per-call timeouts, streaming, and circuit-breaker behavior
- Context compaction (`/compact` command or automatic) with LLM-powered summarization
- Configurable reasoning effort (`/think off|low|medium|high`)
- Delegated sub-agents with configurable profiles, tool restrictions, and depth limits
- Multi-agent routing — route channels/senders with per-route model, prompt, tool preset, and tool allowlist overrides
- Persistent session search, user profiles, and session-scoped todo state available to the agent and operators

### Built-In Providers

- **OpenAI** via `Provider: "openai"`
- **Claude / Anthropic** via `Provider: "anthropic"` or `Provider: "claude"`
- **Gemini** via `Provider: "gemini"` or `Provider: "google"`
- **Azure OpenAI** via `Provider: "azure-openai"`
- **Ollama** via `Provider: "ollama"`
- **OpenAI-compatible** endpoints via `Provider: "openai-compatible"`, `groq`, `together`, or `lmstudio`

OpenClaw registers OpenAI, Claude, and Gemini natively at startup, so a fresh install only needs a provider id, model, and API key to get going.

### Review-First Learning

- The runtime can observe completed sessions and create **pending learning proposals** instead of auto-mutating behavior
- Proposal kinds include **`profile_update`**, **`automation_suggestion`**, and **`skill_draft`**
- Approving a proposal can update a user profile, create a disabled automation draft, or write a managed skill draft and reload skills
- Rejections, approvals, and source-session references are preserved so operators can audit what the system learned and why
- Learning proposals are available over the admin API, `OpenClaw.Client`, and the TUI review flow

### Memory, Profiles, and Automation

- **Session search** spans persisted conversation content for recall and operator lookup
- **User profiles** store stable facts, preferences, projects, tone, and recent intent, with native `profile_read` / `profile_write` tools
- **Automations** support list/get/preview/create/update/pause/resume/run flows and integrate with cron-backed delivery
- **Todos** are persisted per session and available through the native `todo` tool and operator surfaces

### 48 Native Tools

| Category | Tools |
|----------|-------|
| **File & Code** | `shell`, `read_file`, `write_file`, `edit_file`, `apply_patch`, `process`, `git`, `code_exec`, `browser` |
| **Sessions** | `sessions`, `sessions_history`, `sessions_send`, `sessions_spawn`, `sessions_yield`, `session_status`, `session_search`, `agents_list` |
| **Memory** | `memory`, `memory_search`, `memory_get`, `project_memory` |
| **Web & Data** | `web_search`, `web_fetch`, `x_search`, `pdf_read`, `database` |
| **Communication** | `message`, `email`, `inbox_zero`, `calendar` |
| **Home & IoT** | `home_assistant`, `home_assistant_write`, `mqtt`, `mqtt_publish` |
| **Productivity** | `notion`, `notion_write`, `todo`, `automation`, `cron`, `delegate_agent` |
| **Media & AI** | `image_gen`, `vision_analyze`, `text_to_speech` |
| **System** | `gateway`, `profile_read`, `profile_write` |

### Tool Presets & Groups

Named presets control which tools are available per surface:

| Preset | Description |
|--------|-------------|
| `full` | All tools, no restrictions |
| `coding` | File I/O, shell, git, code execution, browser, memory, sessions |
| `messaging` | Message, sessions, memory, profiles, todo |
| `minimal` | `session_status` only |
| `web` / `telegram` / `automation` / `readonly` | Channel-specific defaults |

Presets compose from reusable **tool groups**: `group:runtime`, `group:fs`, `group:sessions`, `group:memory`, `group:web`, `group:automation`, `group:messaging`.

### 9 Channel Adapters

| Channel | Transport | Features |
|---------|-----------|----------|
| **Telegram** | Webhook | Media markers, photo upload, signature validation |
| **Twilio SMS** | Webhook | Rate limiting, opt-out handling |
| **WhatsApp** | Webhook / Bridge | Official Cloud API + Baileys bridge, typing indicators, read receipts |
| **Teams** | Bot Framework | JWT validation, conversation references, group/DM policy |
| **Slack** | Events API | Thread-to-session mapping, slash commands, HMAC-SHA256, mrkdwn conversion |
| **Discord** | Gateway WebSocket | Persistent connection, slash commands, Ed25519 interaction webhook, rate limiting |
| **Signal** | signald / signal-cli | Unix socket or subprocess bridge, privacy mode (no-content logging) |
| **Email** | IMAP/SMTP | MailKit-based |
| **Webhooks** | HTTP POST | Generic webhook triggers with HMAC validation |

All channels support DM policy (`open` / `pairing` / `closed`), sender allowlists, and deduplication.

### Skills

7 bundled skills: `daily-news-digest`, `data-analyst`, `deep-researcher`, `email-triage`, `homeassistant-operator`, `mqtt-operator`, `software-developer`.

Skills are loaded from workspace (`skills/`), global (`~/.openclaw/skills/`), or plugins. Install from ClawHub with `openclaw clawhub install <slug>`.

### Chat Commands

| Command | Description |
|---------|-------------|
| `/status` | Session info (model, turns, tokens) |
| `/new` / `/reset` | Clear conversation history |
| `/model <name>` | Override LLM model for session |
| `/think off\|low\|medium\|high` | Set reasoning effort level |
| `/compact` | Trigger history compaction |
| `/verbose on\|off` | Show tool calls and token usage per turn |
| `/usage` | Show total token counts |
| `/help` | List commands |

### Plugin System

Install community plugins directly from npm/ClawHub:

```bash
openclaw plugins ins

[truncated…]