Litespeed
A local coding agent for your Mac, terminal, and browser, built around multi-model workflows. Use one model or combine a driver with sidekicks, workers, or experts. Follow their work, approve changes, and review the result in one conversation. Pairing models lets a faster, cheaper model handle routine work while a stronger model handles planning or difficult tasks, which can reduce cost and wait time.
Connect through LiteLLM, OpenAI-compatible APIs, native Anthropic, or ChatGPT device sign-in. Your project stays on your machine; prompts and selected context go to your chosen provider.
Quick start
Mac desktop app
Download Litespeed for Apple silicon Macs · macOS 14 or later
Open the downloaded disk image, drag Litespeed onto Applications, then open Litespeed from Applications. The app includes its runtimes and task browser. This is a locally signed desktop preview, not an Apple-notarized release; see the installation instructions if macOS blocks the first launch.
Terminal
Download the terminal package with its runtimes included—no Node or npm setup:
curl -fsSL https://github.com/BerriAI/litespeed/releases/latest/download/install.sh | sh
Open a new terminal, then run it from the project you want to work on:
cd /path/to/your/project
litespeed
The installer sets up the litespeed command for new terminal windows (zsh or Bash). If you use the same terminal window where you installed it, run export PATH="$HOME/.local/bin:$PATH" first.
Litespeed starts its local backend automatically. Updates appear in both UIs, or run litespeed update. Saved sessions and settings stay outside the application package.
See installation, updates, and the source-build alternative. Packages support Apple silicon and Intel Macs. Git and your project's development tools remain separate.
Litespeed has its own command, separate from the LiteLLM gateway CLI. If you used an earlier version of this agent, follow the upgrade guide to carry over saved sessions and configuration.
- On your first launch, connect your LiteLLM gateway base URL and API key, then review your ready-to-use Sidekick Fusion setup with Shunt off. Litespeed picks the latest available driver in this order: Astra → Fable → Opus → Sol; the sidekick uses Sol → Opus → Sonnet. If none are available, it asks you to choose a model. Every setting is editable, including Single model and experimental LiteFusion. Litespeed remembers your choices; running
litespeed in another project opens chat directly. Gateways that do not require a key can leave it blank.
- Use Models or
/models to change your arrangement. Ask first is the default. Handoffs run automatically; choose Allow project edits for routine coding or Allow all tools in permissions or on any routine approval prompt. The full setup is available from the web sidebar or /setup. If you already use Claude Code or Codex skills, /skills (or Settings → Project profiles) offers Import a Claude/Codex skill… to copy one into the project — see Project profiles and skills.
- Type a task. Build can edit files and run commands; Plan uses read-only tools. Type
/ for command suggestions in either client; use ↑/↓, Tab or Enter to complete, and Esc to dismiss. Ctrl+P opens terminal commands and navigation.
See the terminal guide for shortcuts, resuming sessions, and configuration.
Mac desktop app
The desktop build puts projects and tasks on the left, with persistent files, review, browser, and computer views beside the conversation. It can import your existing Litespeed settings and saved tasks into its own local data.
See the desktop app guide for opening the portable app, available workflows, shortcuts, and current limits. Build and packaging details cover the native Mac shell and bundled runtime. The desktop preview is available for Apple silicon Macs.
Prefer the browser?
Once Litespeed is running, open http://localhost:3210. The browser and terminal share saved sessions, providers, and model settings.
For a web-only session, run litespeed serve. For development, use npm run dev from the checkout. See development and updating.
Choose how models work together
| Architecture | How it works |
|---|
| Single model | One model investigates, implements, and checks the task. |
| Sidekick Fusion · Recommended | A strong driver plans and reviews; a cheaper sidekick keeps context across handoffs. |
| LiteFusion · Experimental | One persistent lead routes 63 task categories to specific models and reasoning levels, with one shared hard/escalation map. |
| Team Fusion | A strong driver assigns fresh cheaper workers, runs independent work in parallel, and verifies the combined result. |
| Expert Fusion | A cheaper driver coordinates fresh strong experts and verifies their work. Independent assignments can run in parallel. |
Choose any connected model for each role. An optional Planner model handles Plan mode separately. Litespeed uses your last chosen model arrangement for new sessions across workspaces; existing sessions keep their settings. Cost and quality depend on the models and task. See architecture details and limits.
Additional options
Shunt sends large reads and routine generation to a separate model. Off by default; enable it in Advanced settings during setup or in Models, in either client. How it works and measured results.
Session goals (/goal) keep an objective moving across turns. There is no turn limit unless you set one. Context compaction makes room during long tasks automatically.
Guides
Unrestricted commands run with your local user’s capabilities. Optional workspace command confinement restricts model-initiated commands; permissions alone are not a sandbox. Keep the server local. Provider keys stay server-side, and provider usage may incur charges.
License
Litespeed is licensed under Apache-2.0. Bundled themes and fonts retain their original licenses; see third-party notices.
