Overview

Get from install to a verified first task

  1. Download and install the latest version, following platform availability.
  2. Confirm system requirements and the security defaults.
  3. Register your first Workspace.
  4. Connect at least one Provider and confirm your account status.
  5. Check that thinking, planning, execution, review, and research roles all have somewhere to route.
  6. Complete your first Session.

Browser automation, Remote MCP, and automatic knowledge consolidation stay off by default — only turn them on once you understand their data boundaries. The goal of your first pass is confirming a complete loop across Workspace, Provider, model, routing, and basic tools.

Once your basic Session loop is stable, and you want to reuse the same work definition long-term, talk to a new Bot's own self-builder from Build & refine a Bot. A Bot never inherits the current Session's or build conversation's temporary permissions — you still review and grant its exact capabilities and resource grants yourself.

System requirements#

Platform#

  • macOS 14 or later (Apple Silicon or Intel), Windows (x64 or ARM64), Linux (x86_64 or ARM64).
  • No display in your environment? Use the kxen headless server from a browser — see Web Mode.
  • zsh must be available.
  • Google Chrome, Chromium, or Microsoft Edge is only required if you use the Browser capability.

Project tooling#

  • A Git repository for Worktree, Checkpoint, and the change view.
  • A language server for LSP, matching your language.
  • You supply the runtime for any MCP server yourself.

Models#

You need at least one available Provider. Options include:

  • A supported subscription OAuth login.
  • An API key.
  • Local Ollama.
  • An OpenAI- or Anthropic-compatible endpoint.

Credential availability has to be confirmed through the app's own connection test — a model name appearing in the catalog doesn't mean your current account can actually call it.

Permissions#

The app requests folder, microphone, speech-recognition, or automation permissions only when you actually use the matching capability — you don't need to grant anything ahead of time.

Once these are satisfied, continue to your first Workspace.

Your first Workspace#

Workspace is where kxen connects to your local project. Every Session belongs to a working directory, and project configuration, knowledge, MCP, LSP, Checkpoint, and Worktree are all bounded by it.

Registering a project#

  1. Choose "Add Workspace" in the left sidebar.
  2. Pick your project directory with the system's native directory picker (in Web Mode, type the project's absolute path instead).
  3. Confirm the project shows up in your Workspace list.
  4. Only confirm Workspace trust when you actually trust the repository's content.

What trust affects#

Trust isn't decorative. Only a trusted Workspace loads project configuration, project MCP, project Skills, project Commands, and injectable project knowledge.

An untrusted Workspace can still be browsed, but the project's executable surface and prompt-injection surface stay closed. Personal-scope knowledge isn't affected by a project's trust state either way.

When you're done#

You can see the project in your Workspace list, you can select that directory when creating a Session, and the status bar shows the correct working directory. From there, continue to your first Session.

Your first Session#

Creating one#

  1. Create a Session from your target Workspace.
  2. Confirm the Provider and model in Composer's model picker.
  3. Enter a task with a clear scope and a verifiable result.

A good first task:

Read the current README, and explain the project's startup command and main modules — don't modify any files.

A read-only first task is deliberate: it lets you confirm file reading, model responses, the status bar, and the cancel path.

What to watch while it runs#

  • Current model and context usage.
  • Tool calls and their argument summaries.
  • Approval requests.
  • Background task status.
  • TTFT, elapsed time, output rate, and token usage.

Use Composer's "Stop" button to end the current run. By default, a new message sent while a run is in progress joins that Session's waiting queue.

Verifying the result#

If a follow-up task modifies files:

  1. Review the Session's changes.
  2. Run the project's own check commands.
  3. Confirm there's no unexplained external file change.
  4. Rewind from a specific message if you need to.

Once you've completed this first loop, read Session, Composer, and Context next.

An empty Session's home screen gives you entry points for Goal, Schedule, Workflow, and Team. You can fill in a Goal's objective and completion criteria directly from the right-hand Dock, or let the agent help clarify one with /write-goal; Schedule can be created directly in Settings. Workflow and Team are execution capabilities that need model orchestration — their entry points fill Composer with an explicit template rather than running an opaque script outside the Agent Runtime.

Updated

Was this page helpful?