Open source · Apache 2.0

Run your own
agent.

Your always-on chief of staff.

An AI agent that runs on your hardware, connects to Slack and Discord, and does whatever you delegate to it — research, triage, reporting, the stuff you never get to. Not a coding assistant. Built with .NET. Open source.

netclaw — bash
$ curl -sSL https://releases.netclaw.dev/install.sh | bash
$ netclaw init
$ netclaw daemon start
$ netclaw chat

Up and running in two minutes

curl -sSL https://releases.netclaw.dev/install.sh | bash

Installs CLI + daemon to ~/.netclaw/bin. Supports x64 and arm64.

curl -sSL https://releases.netclaw.dev/install.sh | bash

Same install script. Supports Apple Silicon and Intel Macs.

iwr -useb https://releases.netclaw.dev/install.ps1 | iex

PowerShell. Installs to %LOCALAPPDATA%\Programs\netclaw.

docker run -d --name netclaw -v {your-config-dir}:/root/.netclaw ghcr.io/netclaw-dev/netclaw

Mount your local ~/.netclaw config directory. Configure with NETCLAW_ env vars.

Then
netclaw init

Pick your LLM provider, connect Slack or Discord, set your audience.

Then
netclaw daemon start && netclaw chat

Daemon stays on in the background. Chat from CLI, Slack, or Discord.

Every AI agent you use works for someone else.

Your data trains their models. Your access disappears when they change their pricing.

Netclaw runs on your hardware, connects to your channels, and stays on — 24/7, no exceptions. It handles the stuff you keep putting off: triaging alerts, posting updates, running the checks nobody remembers to run.

Netclaw tool approval prompt in Slack — showing Approve once, Approve for this chat, Approve always, and Deny buttons Click to enlarge

You control what requires approval and what doesn't.

How Netclaw handles a request

Something triggers Netclaw. It figures out what to do, checks whether it needs your OK, then does it and tells you what happened.

Triggers
Webhook
Slack message
Schedule fires
Agent

netclawd

Evaluates context,
picks tools

Decision
Needs
approval?
Action
Yes →

Asks in Slack

No →

Runs automatically

Result

Reports back

Full audit trail
in your channel

1

Event arrives

Webhook, Slack message, or scheduled task

2

netclawd evaluates

Picks the right tools for the job

3

Approval check

Asks in Slack if needed, runs automatically if safe

4

Reports back

Full audit trail in your channel

What's included

Secure by default

Three permission levels — Private, Team, Public — control what each agent can see and do. Default-deny. If a rule is missing, the answer is no.

MCP-first

Native Model Context Protocol support with OAuth. You decide which tools each audience can use and which ones need your approval first.

Automatic memory

Ask it to book a flight once. Next time, it already knows your airline, your home airport, your seat preference. Memory segregated by audience — private stays private.

Small models welcome

Works with models as small as 9B parameters. Netclaw only loads 10-20 tools per call, so small models don't choke even when you have dozens of MCP servers connected.

Webhooks & schedules

Point a webhook at Netclaw and it handles the rest. Cron-style schedules run tasks on their own. You stop being the bottleneck.

Deploy a fleet

Each instance uses ~80MB of RAM. Run one per project, one per client, one per domain — each with its own permissions and memory.

Three pieces, one agent

netclawd

Always-on daemon. LLM sessions, tool execution, persistence. Built on Akka.NET actors.

netclaw

Thin CLI client. Config, management, interactive chat over SignalR.

SkillServer

Self-hosted skill registry. Publish reusable skills and share them between agents.