Product

Clawddesk

An open-source command desk for AI agents, built directly on Anthropic's Claude Agent SDK.

What it is

Clawddesk is a small, hackable multi-agent dashboard built on the Claude Agent SDK. It runs entirely on your own machine: a sidebar of agents, a streaming chat, a durable task queue, a scheduler, and a set of guardrails, all wired to a single SDK call.

It is an educational reference, not a product. There is no hosted version and none is planned. You supply your own Anthropic credentials and run it locally. The point is to be read, studied, and forked.

Built for

Developers evaluating what the Claude Agent SDK makes easy. People who learn best by reading a complete, working example rather than a tutorial. Anyone who wants a personal command desk for a small team of agents, running on their own machine with their own keys.

Why it exists

When Claude is the target model, the Claude Agent SDK collapses the engine layer to a function call. The work that used to take weeks of agent-loop plumbing is now a single query(). Clawddesk is a concrete, readable example of how thin that layer can be, and a sandbox for what the SDK makes easy that used to be hard.

How it is different

  • The SDK is the engine

    Clawddesk does not reimplement the agent loop. It leans on the Claude Agent SDK, where one query() call carries tool use, sub-agents, plan mode, sessions, and hooks. Almost every feature is one or two options on that call.

  • Local and private by default

    There is no hosted version and no account. You bring your own Anthropic credentials and run it on your own machine. Your conversations and keys never leave your computer.

  • Read it, fork it, MIT

    Clawddesk is an educational reference, not a product. The whole thing is open source under the MIT license, written to be studied and modified, so reading the source is reading the SDK’s surface area.

What you can do with it

  • A multi-agent command desk

    Built-in specialists plus unlimited custom agents you spawn yourself, each with its own prompt, tools, and model.

    • A router agent delegates to the right specialist on its own, the model makes the call
    • Token-by-token streaming with the tool-use trace visible as the agent works
    • Per-agent model selection across Opus, Sonnet, and Haiku

  • Run agents unattended

    A durable, SQLite-backed task queue and a cron scheduler let agents keep working on a schedule.

    • Tasks survive a restart, with atomic checkout and lease-based crash recovery
    • Cron runs report to the app, a file, or Telegram, and keep a run history
    • Auto-routing classifies each task to the best agent

  • Guardrails you control

    Dangerous tools pause for your sign-off, and budgets are enforced before every call.

    • Per-task approval gates via a PreToolUse hook
    • Per-agent cost and rate caps checked before each query, so nothing runs away
    • Browser automation sits behind a per-domain allow-list and a private-IP gate

  • Extend every agent

    Context pins, MCP servers, Skills, a Soul Builder personality, a Telegram bridge, and local voice.

    • Per-agent MCP servers over stdio, http, or sse
    • Skills Studio builds, installs, and scans Skills, all on your machine
    • Drive the same agents from your phone over Telegram, or by voice through WhisprDesk

Use cases

  • Learn the Claude Agent SDK

    Read a complete, working dashboard and see exactly how each feature maps to an option on query(). The fastest way to understand how thin the layer has become.

  • A personal command desk

    Run a small team of specialists for your own work, drafting, research, ops, on your own machine, with your own credentials and your own data.

  • A starting point to fork

    Clone it and bend it to your own workflow. It is MIT licensed and written to be modified, not locked down.

  • A teaching example

    A worked reference for anyone learning to build with AI agents, and the developer-facing complement to the free Clawdemy curriculum.

Open source posture

Clawddesk is released under the MIT license, with the full source on GitHub. There is no account, no telemetry, and no cloud component RBJ Global operates. You read the code, clone it, bring your own credentials, and run it on your own machine.

Part of the RBJ Global family

Clawddesk shares the posture that runs across the whole portfolio: local-first, bring-your-own-keys, no telemetry, and a strong preference for keeping data on the user's own machine. It is the developer-facing, educational expression of that idea, applied to the Claude Agent SDK.

Free, open source, no account.

Clone it, bring your own Anthropic credentials, and run it on your own machine. MIT licensed, no cloud, no telemetry.

See the full product site, the interactive architecture diagram, and the documentation:

Visit clawddesk.ai →

Source on github.com/RBJGlobal/clawddesk.

More from RBJ Global

ClaudeLink

Coordinate a team of AI coding agents across terminals.

Read more·claudelink.ai

Clawless Computer

A calm, capable desktop home for AI work.

Read more·clawless.ai

Clawdemy

Free AI education, narrated for life on the move.

Read more·clawdemy.org

WhisprDesk

Local-first transcription. Your voice never leaves your Mac.

Read more·whisprdesk.com

Trading Agents Lab

An open-source lab for studying how multi-agent LLM systems reason about markets.

Read more·tradingagentslab.ai

iLoveMD

Free Markdown viewer and converter. Runs 100% in your browser.

Read more·iluvmd.com