Clawdbot is an open supply private AI assistant that you simply run by yourself {hardware}. It connects giant language fashions from suppliers similar to Anthropic and OpenAI to actual instruments similar to messaging apps, recordsdata, shell, browser and sensible house units, whereas protecting the orchestration layer underneath your management.
The fascinating half is just not that Clawdbot chats. It’s that the undertaking ships a concrete structure for native first brokers, and a typed workflow engine referred to as Lobster that turns mannequin calls into deterministic pipelines.
Structure: Gateway, Nodes and Expertise
On the heart of Clawdbot is the Gateway course of. The Gateway exposes a WebSocket management airplane on ws://127.0.0.1:18789 and a neighborhood HTTP interface for the management UI and internet chat.
Your messages from WhatsApp, Telegram, Sign, Slack, Discord, iMessage and different channels are delivered to the Gateway. The Gateway decides which agent ought to deal with the message, which instruments it could name, and which mannequin supplier to make use of. It then sends the reply again over the identical channel.
The runtime is break up into just a few core ideas:
- Gateway: Routing, mannequin calls, software invocation, classes, presence and scheduling.
- Nodes: Processes that give Clawdbot entry to native sources similar to file system, browser automation, microphone, digital camera or platform particular APIs on macOS, Home windows, Linux, iOS and Android.
- Channels: Integrations for chat programs like WhatsApp, Telegram, Discord, Slack, Sign, Microsoft Groups, Matrix, Zalo and extra. These are configured as channel backends that connect to the Gateway.
- Expertise and plugins: Instruments that the agent can name, described in an ordinary
SKILL.mdformat and distributed by means of ClawdHub.
This separation permits you to run the Gateway on a 5 greenback digital server or a spare machine at house, whereas protecting heavy mannequin compute on distant APIs or native mannequin backends when wanted.
Expertise and the SKILL.md commonplace
Clawdbot makes use of an open expertise format described in SKILL.md. A talent is outlined in Markdown with a small header and an ordered process. For instance, a deployment talent may specify steps similar to checking git standing, working exams and deploying solely after success.
---
title: deploy-production
description: Deploy the present department to manufacturing. Use solely after exams cross.
disable-model-invocation: true
---
1. Test git standing guaranteeing clear working listing.
2. Run `npm check`
3. If exams cross, run `npm run deploy`
The Gateway reads these definitions and exposes them to brokers as instruments with specific capabilities and security constraints. Expertise are revealed to ClawdHub and may be put in or composed into bigger workflows.
Because of this operational runbooks can transfer from ad-hoc wiki pages into machine executable expertise, whereas nonetheless being auditable as textual content.
Lobster: Typed Workflow Runtime for Brokers
Lobster is the workflow runtime that powers Native Lobster and plenty of superior Clawdbot automations. It’s described as a typed workflow shell that lets Clawdbot run multi step software sequences as a single deterministic operation with specific approval gates.
As an alternative of getting the mannequin name many instruments in a loop, Lobster strikes orchestration right into a small area particular runtime:
- Pipelines are outlined as JSON or YAML, or as a compact shell like pipeline string.
- Steps change typed JSON knowledge, not unstructured textual content.
- The runtime enforces timeouts, output limits and sandbox insurance policies.
- Workflows can pause on negative effects and resume later with a
resumeToken.
A easy inbox triage workflow appears to be like like this:
title: inbox-triage
steps:
- id: accumulate
command: inbox record --json
- id: categorize
command: inbox categorize --json
stdin: $accumulate.stdout
- id: approve
command: inbox apply --approve
stdin: $categorize.stdout
approval: required
- id: execute
command: inbox apply --execute
stdin: $categorize.stdout
situation: $approve.accepted
Clawdbot treats this file as a talent. Once you ask it to scrub your inbox, it calls one Lobster pipeline as an alternative of improvising many software calls. The mannequin decides when to run the pipeline and with which parameters, however the pipeline itself stays deterministic and auditable.
Native Lobster is the reference agent that makes use of Lobster to drive native workflows and is described in protection as an open supply agent that redefines private AI by pairing native first workflows with proactive habits.
Proactive native first habits
A key cause Clawdbot is trending and visual on X and in developer communities is that it behaves like an operator, not only a chat window.
As a result of the Gateway can run scheduled jobs and observe state throughout classes, frequent patterns embrace:
- Day by day briefings that summarize calendars, duties and vital mail.
- Periodic recaps similar to weekly shipped work summaries.
- Screens that look ahead to situations, then message you first in your most well-liked channel.
- File and repository automations that run domestically however are triggered by pure language.
All of this runs with routing and power coverage in your machine or server. Mannequin calls nonetheless go to suppliers like Anthropic, OpenAI, Google, xAI or native backends, however the assistant mind, reminiscence and integrations are underneath your management.
Set up and developer workflow
The undertaking supplies a one line installer that fetches a script from clawd.bot and bootstraps Node, the Gateway and core elements. For extra management, you’ll be able to set up through npm or clone the TypeScript repository and construct with pnpm.
Typical steps:
curl -fsSL https://clawd.bot/set up.sh | bash
# or
npm i -g clawdbot
clawdbot onboard
After onboarding you join a channel similar to Telegram or WhatsApp, select a mannequin supplier and allow expertise. From there you’ll be able to write your individual SKILL.md recordsdata, construct Lobster workflows and expose them by means of chat, internet chat or the macOS companion software.
Some Examples
Michal Sutter is a knowledge science skilled with a Grasp of Science in Knowledge Science from the College of Padova. With a stable basis in statistical evaluation, machine studying, and knowledge engineering, Michal excels at remodeling advanced datasets into actionable insights.
