clawdotnet/openclaw.net

Self-hosted Enterprise and Personal AI + agent runtime in .NET (NativeAOT-friendly)

C#

518

1,129 commits

updated Oct 3, 2026

See the code

README

OpenClaw.NET logo

OpenClaw.NET

A self-hosted AI agent runtime for .NET, with a desktop companion for everyday use.

CI License: MIT NativeAOT-friendly Ask DeepWiki

Download desktop · Run from source · Documentation · Compatibility · 中文

Run an assistant locally or host a gateway for your team. Connect a hosted or local model, give it tools and memory, and work through desktop chat, the browser, messaging channels, or APIs. Developers can extend the runtime in .NET and deploy supported capabilities with NativeAOT.

OpenClaw.NET is the runtime and repository. AgentQi Companion is the Avalonia desktop product for OpenClaw.NET; AgentQi.dev is the documentation and ecosystem home.

This is an independent .NET implementation inspired by OpenClaw, with practical ecosystem compatibility. It is not affiliated with or endorsed by the upstream project.

Meet the desktop companion

Chat is the starting point. AgentQi Companion brings setup, connections, activity, memory, history, and settings into one desktop interface, with light and dark themes.

AgentQi Companion for OpenClaw.NET: chat welcome screen, setup shortcuts, navigation sidebar, and Configure via chat option

Companion welcome screen from main, before connecting a gateway. Packaged releases may show an earlier interface.

  • Set up through the app: choose a provider and model, enter credentials in dedicated fields, and start a local gateway.
  • Describe a settings change: use /configure set the session timeout to 45 minutes, review the proposed values, then apply or cancel.
  • Use the GUI when you prefer: the manual settings editor works without a language model when the gateway is reachable; advanced setup is available when needed.

Chat configuration covers supported scalar settings, not every operation. Credentials, provider setup, and complex account objects retain dedicated forms. See what can be configured through chat.

Download And Run Desktop

Choose a desktop bundle from the latest release. Each includes AgentQi Companion, the NativeAOT gateway, and the NativeAOT CLI.

  1. Extract the archive and launch Companion from the companion folder.
  2. Open Setup. In the redesigned Companion, you can also choose Set up my assistant in Chat.
  3. Choose your provider and model. Enter a hosted provider key, or choose Ollama for a local model server.
  4. Click Set Up and Start.

Companion writes a local configuration, starts the bundled gateway on 127.0.0.1, and connects to it. Embedded local inference is also available; see the local model guide for package installation and requirements.

The Windows and macOS release archives are currently unsigned. See the release guide for first-run OS warnings, checksums, standalone CLI/gateway downloads, and signing status. The screenshot above shows development on main; check release notes for the features included in your download.

Quickstart

Run a local gateway from source

Install Git and the .NET 10 SDK, then clone the repository:

git clone https://github.com/clawdotnet/openclaw.net
cd openclaw.net

For the default hosted-provider path, set your API key and start the guided setup. These commands use Bash or zsh:

export MODEL_PROVIDER_KEY="sk-..."
dotnet run --project src/OpenClaw.Cli -c Release -- start

In PowerShell, set the key with $env:MODEL_PROVIDER_KEY = "sk-...", then run the same dotnet command. To use another provider or a local model, follow the provider setup guide.

The CLI runs setup if needed, saves the local configuration, and launches the gateway. Wait for OpenClaw gateway ready., then open:

SurfaceAddress
Browser chathttp://127.0.0.1:18789/chat
Administrationhttp://127.0.0.1:18789/admin
Integration status APIhttp://127.0.0.1:18789/api/integration/status
MCP endpointhttp://127.0.0.1:18789/mcp

Use named operator accounts for browser administration and operator account tokens for Companion, CLI, API, and WebSocket clients. The quickstart covers the full first-run walkthrough and troubleshooting.

To try the Avalonia Companion from the same checkout:

dotnet run --project src/OpenClaw.Companion -c Release

Connect it to your running gateway through Setup & runtime. See Companion token storage for authentication and credential handling.

Try the runtime without an API key

From the repository root, this deterministic sample demonstrates the runtime loop and tool invocation without a model server, Docker, or a browser:

dotnet restore OpenClaw.Net.slnx
dotnet build OpenClaw.Net.slnx --configuration Release --no-restore
dotnet run --project samples/OpenClaw.HelloAgent -c Release --no-build

Expected output:

OpenClaw.HelloAgent
User: hello
Agent: hello from OpenClaw.NET
Tool: echo(hello): ok

What Works Now

AreaCapabilities
Agent runtimeStreaming, tool execution, cancellation, retries, sessions, memory, and token-usage reporting.
InterfacesAgentQi Companion, browser chat and admin UI, CLI, terminal UI, OpenAI-compatible HTTP endpoints, MCP, and WebSockets.
ModelsOpenAI, Claude, Gemini, Azure OpenAI, DeepSeek, Ollama, and OpenAI-compatible providers; named profiles and optional embedded local inference.
Decision routingOpt-in hosted Jev or local Laya routing, with shadow evaluation, safety floors, and fallback to the existing policy. Laya active routing requires a calibration artifact.
Tools and channels80+ native and optional tool surfaces for files, web, sessions, databases, email, home automation, and more; adapters for Telegram, WhatsApp, Teams, Slack, Discord, and other channels. The active set depends on configuration.
ExtensionsMCP servers and interactive MCP Apps, reusable SKILL.md packages, first-party .NET integrations, and supported OpenClaw TS/JS plugins.
MetaSkillsReusable, auditable DAG workflows that compose skills, tools, and LLM calls with dependency-aware scheduling, parallel steps, fallbacks, and user-input checkpoints. See the MetaSkills guide.
MetaSkill APIDirect invocation of a named MetaSkill DAG in an existing session, with durable idempotency and result replay. See the invocation API.
Long-running workSession-scoped /goal continuation, /loop recurring prompts, and optional durable workflow backends.
Structured project memoryOptional Fractal Memory capture, review, import, handoff, and resume workflows through agent tools, CLI, and admin API. Writes are opt-in and agent approvals are enabled by default.
Review and observabilityTool approvals, diagnostics, audit and trajectory exports, passive harness contracts, evidence bundles, and optional Plan-Execute-Verify execution.
Developer toolingOffline harness regression checks, static codebase maps, SkillKit authoring and validation, and TokenJuice tool-output reduction.

Explore the user guide, tool catalog, SkillKit, and harness testing for details. The separate AgentQi Mobile repository provides an Android operator console and build instructions.

For audience restrictions, verified bundle updates, remote device enrollment, and a one-command local demo, see the operations guide. Audience controls are opt-in; verified updates require an independently trusted publisher key, and remote device enrollment requires HTTPS and an administrator-issued code.

Compatibility and capability boundaries

NativeAOT support and upstream compatibility depend on the feature you enable. Check the capability matrix and compatibility guide before choosing a deployment lane.

LaneExamples
CoreRuntime loop, gateway, CLI, OpenAI-compatible API, and NativeAOT-friendly host path.
OptionalCompanion, channel adapters, browser/MQTT packages, model integrations, and workflow backends.
ExperimentalEmbedded local model sidecars and adapter-oriented package paths.
JIT-onlyDynamically loaded native .NET plugins.

For framework and agent interoperability, see Microsoft Agent Framework, A2A, and workflow backends. For direct MetaSkill API calls, see the MetaSkill invocation API or its 简体中文版. For interactive MCP App hosting, use the documented gateway host routes.

Security

Local setup binds the gateway to loopback. Before exposing it beyond your machine, read SECURITY.md: the gateway refuses unsafe public-bind configurations until required authentication and tooling restrictions are in place.

Outbound web fetches and browser navigations use URL safety checks by default, including restrictions on private, loopback, link-local, and metadata destinations. See the security guide before changing those policies.

For private access while keeping the gateway on 127.0.0.1, use Tailscale Serve:

openclaw setup tailscale serve

This command assumes the CLI is on your PATH; from a checkout, use dotnet run --project src/OpenClaw.Cli -c Release -- setup tailscale serve.

Docs

Browse AgentQi.dev or the complete repository documentation index.

Start hereWhat you will find
Evaluator overviewWhat the project does and how to evaluate it.
QuickstartFirst-run setup, provider examples, diagnostics, and recovery.
Getting startedRepository map, architecture, and contributor setup.
User guideProviders, tools, skills, memory, channels, and daily operation.
Companion chat configurationSupported settings, review/apply behavior, and advanced setup.
ReleasesDownloads, checksums, signing status, and release procedures.
Capability matrixCore, optional, experimental, and JIT-only features.
Local modelsEmbedded model packages, sidecars, and frame-based video support.
Capability resolutionLocal discovery, optional Nacos adapters, caches, and replay.
Architecture boundariesRuntime, gateway, extension, and AOT/JIT boundaries.
RoadmapPlanned work and current priorities.
中文文档Simplified Chinese first-run orientation.

Contributing

Contributions are welcome, especially security reviews, NativeAOT improvements, channel adapters, and performance benchmarks. Start with CONTRIBUTING.md.

See project governance, maintainer roles, and the review checklist for contribution and review expectations.

License

MIT

agent-harness
agentqi
agent-runtime
ai-agent
automation
csharp
jev-model
llm
loop-engineering
memory
microsoft-agent-framework
nativeaot
openai-compatible-api
personal-ai
self-evolving
self-evolving-agents
self-hosted
tool-calling
tool-execution

Significant stargazers

maikebing

243 followers · starred Jun 2026

Burak Bayır

138 followers · starred May 2026

Magian

17 followers · starred Apr 2026

clawdotnet/openclaw.net

Self-hosted Enterprise and Personal AI + agent runtime in .NET (NativeAOT-friendly)

C#

518

1,129 commits

updated Oct 3, 2026

See the code

README

OpenClaw.NET logo

OpenClaw.NET

A self-hosted AI agent runtime for .NET, with a desktop companion for everyday use.

CI License: MIT NativeAOT-friendly Ask DeepWiki

Download desktop · Run from source · Documentation · Compatibility · 中文

Run an assistant locally or host a gateway for your team. Connect a hosted or local model, give it tools and memory, and work through desktop chat, the browser, messaging channels, or APIs. Developers can extend the runtime in .NET and deploy supported capabilities with NativeAOT.

OpenClaw.NET is the runtime and repository. AgentQi Companion is the Avalonia desktop product for OpenClaw.NET; AgentQi.dev is the documentation and ecosystem home.

This is an independent .NET implementation inspired by OpenClaw, with practical ecosystem compatibility. It is not affiliated with or endorsed by the upstream project.

Meet the desktop companion

Chat is the starting point. AgentQi Companion brings setup, connections, activity, memory, history, and settings into one desktop interface, with light and dark themes.

AgentQi Companion for OpenClaw.NET: chat welcome screen, setup shortcuts, navigation sidebar, and Configure via chat option

Companion welcome screen from main, before connecting a gateway. Packaged releases may show an earlier interface.

  • Set up through the app: choose a provider and model, enter credentials in dedicated fields, and start a local gateway.
  • Describe a settings change: use /configure set the session timeout to 45 minutes, review the proposed values, then apply or cancel.
  • Use the GUI when you prefer: the manual settings editor works without a language model when the gateway is reachable; advanced setup is available when needed.

Chat configuration covers supported scalar settings, not every operation. Credentials, provider setup, and complex account objects retain dedicated forms. See what can be configured through chat.

Download And Run Desktop

Choose a desktop bundle from the latest release. Each includes AgentQi Companion, the NativeAOT gateway, and the NativeAOT CLI.

  1. Extract the archive and launch Companion from the companion folder.
  2. Open Setup. In the redesigned Companion, you can also choose Set up my assistant in Chat.
  3. Choose your provider and model. Enter a hosted provider key, or choose Ollama for a local model server.
  4. Click Set Up and Start.

Companion writes a local configuration, starts the bundled gateway on 127.0.0.1, and connects to it. Embedded local inference is also available; see the local model guide for package installation and requirements.

The Windows and macOS release archives are currently unsigned. See the release guide for first-run OS warnings, checksums, standalone CLI/gateway downloads, and signing status. The screenshot above shows development on main; check release notes for the features included in your download.

Quickstart

Run a local gateway from source

Install Git and the .NET 10 SDK, then clone the repository:

git clone https://github.com/clawdotnet/openclaw.net
cd openclaw.net

For the default hosted-provider path, set your API key and start the guided setup. These commands use Bash or zsh:

export MODEL_PROVIDER_KEY="sk-..."
dotnet run --project src/OpenClaw.Cli -c Release -- start

In PowerShell, set the key with $env:MODEL_PROVIDER_KEY = "sk-...", then run the same dotnet command. To use another provider or a local model, follow the provider setup guide.

The CLI runs setup if needed, saves the local configuration, and launches the gateway. Wait for OpenClaw gateway ready., then open:

SurfaceAddress
Browser chathttp://127.0.0.1:18789/chat
Administrationhttp://127.0.0.1:18789/admin
Integration status APIhttp://127.0.0.1:18789/api/integration/status
MCP endpointhttp://127.0.0.1:18789/mcp

Use named operator accounts for browser administration and operator account tokens for Companion, CLI, API, and WebSocket clients. The quickstart covers the full first-run walkthrough and troubleshooting.

To try the Avalonia Companion from the same checkout:

dotnet run --project src/OpenClaw.Companion -c Release

Connect it to your running gateway through Setup & runtime. See Companion token storage for authentication and credential handling.

Try the runtime without an API key

From the repository root, this deterministic sample demonstrates the runtime loop and tool invocation without a model server, Docker, or a browser:

dotnet restore OpenClaw.Net.slnx
dotnet build OpenClaw.Net.slnx --configuration Release --no-restore
dotnet run --project samples/OpenClaw.HelloAgent -c Release --no-build

Expected output:

OpenClaw.HelloAgent
User: hello
Agent: hello from OpenClaw.NET
Tool: echo(hello): ok

What Works Now

AreaCapabilities
Agent runtimeStreaming, tool execution, cancellation, retries, sessions, memory, and token-usage reporting.
InterfacesAgentQi Companion, browser chat and admin UI, CLI, terminal UI, OpenAI-compatible HTTP endpoints, MCP, and WebSockets.
ModelsOpenAI, Claude, Gemini, Azure OpenAI, DeepSeek, Ollama, and OpenAI-compatible providers; named profiles and optional embedded local inference.
Decision routingOpt-in hosted Jev or local Laya routing, with shadow evaluation, safety floors, and fallback to the existing policy. Laya active routing requires a calibration artifact.
Tools and channels80+ native and optional tool surfaces for files, web, sessions, databases, email, home automation, and more; adapters for Telegram, WhatsApp, Teams, Slack, Discord, and other channels. The active set depends on configuration.
ExtensionsMCP servers and interactive MCP Apps, reusable SKILL.md packages, first-party .NET integrations, and supported OpenClaw TS/JS plugins.
MetaSkillsReusable, auditable DAG workflows that compose skills, tools, and LLM calls with dependency-aware scheduling, parallel steps, fallbacks, and user-input checkpoints. See the MetaSkills guide.
MetaSkill APIDirect invocation of a named MetaSkill DAG in an existing session, with durable idempotency and result replay. See the invocation API.
Long-running workSession-scoped /goal continuation, /loop recurring prompts, and optional durable workflow backends.
Structured project memoryOptional Fractal Memory capture, review, import, handoff, and resume workflows through agent tools, CLI, and admin API. Writes are opt-in and agent approvals are enabled by default.
Review and observabilityTool approvals, diagnostics, audit and trajectory exports, passive harness contracts, evidence bundles, and optional Plan-Execute-Verify execution.
Developer toolingOffline harness regression checks, static codebase maps, SkillKit authoring and validation, and TokenJuice tool-output reduction.

Explore the user guide, tool catalog, SkillKit, and harness testing for details. The separate AgentQi Mobile repository provides an Android operator console and build instructions.

For audience restrictions, verified bundle updates, remote device enrollment, and a one-command local demo, see the operations guide. Audience controls are opt-in; verified updates require an independently trusted publisher key, and remote device enrollment requires HTTPS and an administrator-issued code.

Compatibility and capability boundaries

NativeAOT support and upstream compatibility depend on the feature you enable. Check the capability matrix and compatibility guide before choosing a deployment lane.

LaneExamples
CoreRuntime loop, gateway, CLI, OpenAI-compatible API, and NativeAOT-friendly host path.
OptionalCompanion, channel adapters, browser/MQTT packages, model integrations, and workflow backends.
ExperimentalEmbedded local model sidecars and adapter-oriented package paths.
JIT-onlyDynamically loaded native .NET plugins.

For framework and agent interoperability, see Microsoft Agent Framework, A2A, and workflow backends. For direct MetaSkill API calls, see the MetaSkill invocation API or its 简体中文版. For interactive MCP App hosting, use the documented gateway host routes.

Security

Local setup binds the gateway to loopback. Before exposing it beyond your machine, read SECURITY.md: the gateway refuses unsafe public-bind configurations until required authentication and tooling restrictions are in place.

Outbound web fetches and browser navigations use URL safety checks by default, including restrictions on private, loopback, link-local, and metadata destinations. See the security guide before changing those policies.

For private access while keeping the gateway on 127.0.0.1, use Tailscale Serve:

openclaw setup tailscale serve

This command assumes the CLI is on your PATH; from a checkout, use dotnet run --project src/OpenClaw.Cli -c Release -- setup tailscale serve.

Docs

Browse AgentQi.dev or the complete repository documentation index.

Start hereWhat you will find
Evaluator overviewWhat the project does and how to evaluate it.
QuickstartFirst-run setup, provider examples, diagnostics, and recovery.
Getting startedRepository map, architecture, and contributor setup.
User guideProviders, tools, skills, memory, channels, and daily operation.
Companion chat configurationSupported settings, review/apply behavior, and advanced setup.
ReleasesDownloads, checksums, signing status, and release procedures.
Capability matrixCore, optional, experimental, and JIT-only features.
Local modelsEmbedded model packages, sidecars, and frame-based video support.
Capability resolutionLocal discovery, optional Nacos adapters, caches, and replay.
Architecture boundariesRuntime, gateway, extension, and AOT/JIT boundaries.
RoadmapPlanned work and current priorities.
中文文档Simplified Chinese first-run orientation.

Contributing

Contributions are welcome, especially security reviews, NativeAOT improvements, channel adapters, and performance benchmarks. Start with CONTRIBUTING.md.

See project governance, maintainer roles, and the review checklist for contribution and review expectations.

License

MIT

agent-harness
agentqi
agent-runtime
ai-agent
automation
csharp
jev-model
llm
loop-engineering
memory
microsoft-agent-framework
nativeaot
openai-compatible-api
personal-ai
self-evolving
self-evolving-agents
self-hosted
tool-calling
tool-execution

Significant stargazers

maikebing

243 followers · starred Jun 2026

Burak Bayır

138 followers · starred May 2026

Magian

17 followers · starred Apr 2026

Languages

C#

88.7%

HTML

5.8%

JavaScript

2.1%

Python

1.6%