What each screen is for, what to do there, and what outcome to expect.

Home

The dashboard you land on. Glance here to see if everything's running, then jump to wherever you need to go.

Operational front door for instance health, routing status, host pressure, active work, and high-value jumps into the rest of the app.

• Check whether any instances are ready for requests.
• Look for machine pressure before starting more models.
• Jump to Runtime, API Access, Chat, or diagnostics when something is not ready.

Download

Where you grab new stuff before using it: AI models, plug-in tools, and skills.

Acquire capabilities before you run them.

• HuggingFace: search, sort, filter, inspect model cards, and download MLX or GGUF models.
• ClawHub: browse skills and plugins, inspect trust signals, install packages, repair, update, and uninstall.
• MCP: browse official and Smithery registry entries, install local/hosted servers, and configure transport.
• Skills: install and manage reusable behavior bundles.

Runtime

The engine room. This is where downloaded models actually start running so you can talk to them.

Where model assets become running API-capable instances.

• Instances: create, start, stop, edit, diagnose, and tune runtime instances.
• Models: inspect installed model assets, size, quantization, fit, loaded state, dependent instances, Finder path, cache, and delete actions.

Chat

The main place to talk to the AI. Type, attach files, get answers — like ChatGPT but on your machine.

Primary conversation workspace with streaming, attachments, tool calls, approvals, session controls, slash commands, compaction, exports, and model recommendations.

• Use it for normal work, prompt testing, file analysis, multimodal inputs, and guided tool use.
• Use /commands to see every built-in and custom slash command.
• Use /compact or auto-compaction for long sessions.

Assistant

A friendlier "always there" version of the AI you can pop out into a small floating window for quick help.

Dedicated in-app assistant surface. Can detach into a floating desktop presence for ongoing companion-style assistance — not intended for carefully audited multi-agent runs.

Work Modes

Saved setups for specific kinds of work. Pick a Work Mode and the app pre-configures the model, memory, tools, and starter agents for that activity.

Preset profiles bundling model family, execution mode, memory behavior, autonomy level, web tools, starter agents/workspaces/automations, and package requirements.

• Create, clone, edit, delete, apply, and reapply saved modes.
• Use mode health to find missing packages or blocked requirements.
• Use package overrides when a saved mode needs custom marketplace behavior.

Knowledge

Everything the AI can "know" about you and your work — facts it remembers, files it can search, notes you've written, even people in your contacts.

The context layer chats and agents can draw from.

• AI Context: active system prompts and shared prompt shaping.
• Skills: installed and draft skill bundles.
• Memory: persistent facts, manual entries, auto-extraction, markdown export, delete controls.
• Documents: RAG files, folder indexing, chunking, search, selected embedding instance.
• Wiki: curated pages, links, backlinks, tags, edit history.
• Dreaming: idle-generated diary entries, memory-palace pages, and insights.
• Contacts: people, groups, tags, and reachability context.
• Obsidian: vault connection, note search, and import.

Workspaces

A team of AI agents working together on a job. You set up the team once and run it whenever you need.

Repeatable multi-agent work.

• Directory: create and edit agents and workspace definitions.
• Run: select a workspace, launch a task, monitor messages, resolve tool confirmations, and stop runs.
• Agents can have roles, models, ports, skills, memory scopes, tool access modes, standing orders, hooks, and prompt previews.

Automation

Schedule the AI to run jobs by itself — on a timer, when something happens, or on a regular heartbeat.

Converts reliable prompts, agents, and tools into scheduled or event-driven work.

• Flows: tasks, hooks, orders, advanced jobs, heartbeat, background jobs, workflows, and cron editor.
• Triggers: inbound webhook endpoints.
• Use approvals for any automation that can write files, run commands, use browsers, call channels, or affect external systems.

Extensions

Add-ons. Connect the AI to outside services like Slack or Telegram, manage installed plug-ins, or hook in extra tools.

Installed capabilities and external surfaces.

• Installed: package runtime health, setup, update, repair, and uninstall.
• Channels: shared inbox, Telegram, WebChat, auto-reply, DM Access, broadcast, accounts, and locations.
• Plugins: app plugin lifecycle and contributed capabilities.
• MCP: server configuration and discovered tools/resources.
• Skills: skill bundle management.

System

Settings, logs, and the "what's going on under the hood?" view. Where you go when something isn't working.

Admin and observability center.

• Settings: core app mode, appearance, storage, API reference, and maintenance.
• Requests: traffic log and request inspector.
• Logs: runtime, app, and crash log history.
• Health: diagnostics and doctor checks.
• Sessions: session manager and archives.
• Usage: request, token, and cost trends.
• Advanced: security, connectivity, media/device, and operator tools.

Lab

A sandbox to try prompts and compare models before committing them to a real workflow.

Where you test before depending on a setup.

• Playground: raw API request testing.
• Compare: side-by-side prompt evaluation.
• Benchmark: TTFT, throughput, and latency checks.
• Browser: managed browser sessions, profiles, previews, and LLM action log.
• Canvas: live HTML, charts, interactive outputs, source, snapshots, reset.
• QA Lab: test suites and regression probes.