Welcome to Larkium.

01 / Open your workspace

Get an available build from the download section, or follow the source build instructions. Native macOS Apple Silicon previews have local validation; public signed binaries and Linux/Windows release validation remain pending.

Open your project folder. New Larkium profiles show chat at startup. You can reopen it from the Larkium Chat status-bar button, the title-bar action or View → Open Larkium Chat. Cmd L on macOS or Ctrl L also opens chat.

02 / Choose a model

Open chat Settings and add a model. Claude, OpenAI-compatible providers and local Ollama servers are supported. Use your provider’s model ID and endpoint; local servers must already be running.

Set the API key when asked, or through model settings. Keys are stored in the OS keychain rather than project settings. Linux needs a Secret Service keyring. A local model without authentication may not need a key.

Choose a tool-capable model for agents that inspect and change code. Use the model check action before relying on tool calling. Provider fees and local hardware requirements depend on the model you choose.

03 / Add the right context

Use @ in chat to mention files, or add a selection from the editor or terminal. Keyword and syntax-aware codebase search are available; semantic search requires an embedding model you configure.

Start with a focused request: “Find where this request is handled, explain the current behavior, and suggest a change.” Review command and edit approvals as they arrive.

04 / Run a command and inspect the result

Ask for a completed shell command block. Click Run in terminal, choose the workspace in a multi-root project, and approve it under your command rules. The configured terminal shell runs the command.

Give me a sh block containing:
printf 'terminal demo\n'; pwd
Do not run it with your tools.
After I click Run in terminal, explain the result.

The result card retains output and exit status. Supported test runners show success rates, failed names and failure output; select Tests for a custom runner. The originating agent follows up on a known completed result.

Shell integration is required for captured execution. If capture is unavailable, Larkium explains the manual fallback rather than silently sending a command. Stop targets the selected chat. An interrupted capture never claims success.

05 / Review a specialist’s changes

Define a specialist in .larkium/agents/*.md, with instructions and explicit tools. Read-only agents can investigate; writing agents opt into isolated-edit execution and work in their own Git worktrees.

Inspect the proposed diff and test evidence, then choose Apply reviewed changes. Larkium checks for changes in your primary checkout and refuses stale proposals. Conflicts require explicit resolution and a refreshed review. Durable checkpoints support Undo.

A worktree isolates repository files; commands still run with your account’s host access. Keep permissions narrow and review the evidence before applying.

06 / Keep track of background work

Switch chats while submitted agents work. Runs, approvals and output stay attached to their own conversation. History distinguishes running, waiting, unread completions and unseen failures.

Scroll up without being pulled back to new text. The ↓ New response control returns to the latest output. After an editor restart, unfinished runs are marked interrupted; they do not silently resume.

07 / Extend your editor

Use Open VSX for editor extensions. Larkium plugins can provide instruction components such as agents, rules and skills, plus supported MCP declarations. Git and local installations are captured as immutable snapshots.

Installation is inert. Enable a plugin with explicit local approval; changed snapshots require approval again. Imported Cursor/Claude bundles use a supported subset rather than their complete runtimes.