Klaus Desktop

Your agent workspace

Give it the outcome. It plans first and waits for your approval, then edits files, runs commands and verifies the result. Review the changes file by file, and confirm the moves that matter.

Supports Apple Silicon Macs (macOS 13 or later) and Windows x64 (Windows 10 or later)

Plan and approval

A broad goal becomes a sequence of steps you can act on, with the reasoning and status of each in the Overview panel. Change one midway and carry on.

How plan mode works

Review changes file by file

The diff view starts with per-file counts, then opens a single file to show the actual lines. Once it looks right, commit the whole repo in one go.

Workspaces and Git

Keeps running when you're away

Checks, daily reports and information gathering run on a schedule, with results sent to Feishu, DingTalk or WeChat — the places you already use.

Scheduled tasks

New: friend collaboration

Add a friend by email, @ them in an ordinary task, and let your agents work it together. The person receiving it decides which directories, apps and execution mode to authorise.

About friend collaboration

How one task runs from start to finish

Three moments from the same task in Klaus, all taken from real screen recordings.

  1. 01

    Planning and execution

    A broad goal is broken into a plan, and tools are called one at a time as it runs. Add a line at any point to change direction — no need to wait for it to stop.

    Klaus desktop main view: session list on the left, transcript and plan card in the middle, Overview panel on the right
  2. 02

    Changes and review

    Expand a sub-agent's progress and changes in the Overview panel: per-file counts first, then the actual edits file by file.

    The Klaus Overview panel: change counts and the per-file diff
  3. 03

    Schedules and reports

    Routine tasks run on their own schedule, the run history stays in the scheduled-tasks panel, and the conclusions go to the chat channels you already use.

    The Klaus scheduled-tasks panel: the schedule list and its run history

How Klaus works

A good tool should widen what a person can do.

The best model is the one you choose

Plug in your own Claude, Codex, Qoder or Kimi subscription, or any OpenAI-compatible endpoint. Configure the high, medium and low effort levels separately and pick the right one for the task in front of you.

Model sources and switching

Let your agents work a task together

Find a friend by their registered email, start a collaboration from an ordinary task, and leave the choice of apps, directories and execution mode to the person receiving it.

Sample accounts and tasks are shown. Friend collaboration needs Klaus desktop 0.4.28 or later on both sides. You can remember an authorisation so that friend's later tasks start automatically on this computer within the scope you chose; revoking it stops tasks already authorised. Removing a friend ends the relationship on both sides and stops authorised tasks, while history is kept.

and more

Each one has a doc page explaining what it does and where to set it up.

Memory

Turn it on and your project conventions, release process and preferences accumulate as you work. Every entry can be viewed, edited and deleted.

View docs

Skills and plugins

Write a routine you repeat as a skill and reuse it in one sentence. Plugins and MCP tools are picked up per task.

View docs

Sub-agents

Split a large task across sub-agents running in parallel, with their plans, progress and artifacts in the Overview panel.

View docs

Parallel worktrees

Changes land on separate branches that don't interfere, so several tasks can run at once — and you decide what gets merged.

View docs

Spec mode

When the requirement isn't settled yet, write the spec document together first and start implementing once it's final.

View docs

Voice input

Hold to talk and dictate the requirement or an extra note; it turns into text and sends.

View docs

What you probably want to know before installing

How do I start?

Install it, then sign in with a Klaus account or continue as a guest. Create a project from a real project directory and say clearly what outcome you want. Its file operations and commands happen inside that directory, and complex work produces a plan for your approval first.

Which systems are supported?

The desktop app currently supports Apple Silicon Macs (macOS 13 or later) and Windows x64 on Windows 10 or later. the macOS installer is available, and the Windows installer is released.

Which models does it use?

Two kinds of source: your own Claude Code, Codex, Qoder, QwenWork, QoderWork, WorkBuddy or Kimi subscription, or a custom API (an OpenAI-compatible endpoint with your own base URL and model name). It runs on your account and your quota — Klaus neither meters nor reports it. When a model is called, the task content that call needs goes to the service you chose.

Can the important actions run away?

Before writing a file, running a command or touching an external service, a permission card appears: allow once, allow for this session, or always allow. A deny rule you set overrides every mode. Scheduled tasks run automatically and don't wait for per-item approval, so agree the task's scope and permission rules before setting one up.

Need help, or want to give feedback?

The docs cover features and internals. The in-app feedback entry attaches the session and environment details automatically, or write to support@klaus-ai.cn.

Hand your next real task to Klaus