Eta

A general-purpose agent for your desktop.

English · 简体中文

Eta is a desktop agent for working with files, understanding images, and turning natural-language instructions into actions. Our goal is to make one agent useful across everyday tasks, from exploring a codebase to organizing a workspace.

What you can do

• Choose your model. Connect an AI account or bring your own API key.
• Give it context. Work in a local project and attach images for vision-capable models.
• Let it act. Read, write, and edit files or run commands. Enable only the tools you want.
• Keep your work. Return to saved conversations and continue where you left off.

Get started

1. Connect a model in Settings.
2. Start a new chat and select a project folder.
3. Describe your task. Attach pictures through @image-path, or paste and drag them in.

Use Settings → Permissions to control image reading and Settings → Tools to enable or disable built-in tools. External tools are not available yet.

Keep chats, settings, and web pages open in the top tab bar. The + opens a web tab; enter an address or search in its address bar. New chat returns to an existing unsent chat, preserving its draft and project selection. After submitting it, you can start another chat. Closing a chat tab keeps its saved history and lets an admitted task continue. Use Cmd/Ctrl+W to close a tab and Cmd/Ctrl+Shift+T to reopen it. Switching tabs preserves each chat's draft and project selection.

Use Settings → Application → Storage to check session usage and search or sort local sessions. Click a session to locate its folder in your file manager. The page also provides paths to the sessions directory, model settings, and Catalog.

Right-click a saved chat to rename it, change its project grouping, archive or restore it, open it in another Eta window, or permanently delete its local history. Project grouping does not change the execution directory. Open with opens the main JSONL transcript in the default or a selected application; use a text editor to inspect it.

Skills

Place skill folders containing SKILL.md in ~/.agents/skills/ for personal use or .agents/skills/ in a project. Use Settings → Skills to view and open skill directories, add custom directories, inspect discovery errors, and enable or disable skills. Opening a missing skill directory creates it. Project skills take precedence when names collide.

Eta first offers the agent skill names and descriptions. The agent loads instructions when relevant and reads supporting files only as needed. Choose a skill from the composer, use Cmd/Ctrl+Shift+K, or include $skill-name in your request to select it explicitly. The conversation retains activated instructions across restarts and context compaction. Remove a loaded skill in the composer before loading an updated version. Skills use the tools and dependencies already available in your environment.

Subagents

Open the right sidebar from the top-right window control (Alt+←) to inspect delegated work. Click a task to open its conversation and view reported progress. Send a message while it is working to adjust its direction or request a summary of existing findings; sending to a settled or paused task continues the same conversation. Independent children receive their assigned task and project instructions; forked children inherit the conversation history at creation.

Use Settings → Subagents to create presets with instructions, thinking levels, and candidate models. Enable further delegation only for presets that coordinate other agents; workers complete their assigned task directly by default. The model scope follows the current main model by default; enable other models in the model picker. Set nesting depth and concurrency per chat. Opportunistic mode delegates when useful. Orchestrator mode delegates tool execution while the main conversation handles discussion, coordination, and results.

In orchestrator mode the main conversation remains available between coordinating turns while children keep working. Completion events bring it back to summarize their results. A wait timeout does not stop the child. Stopping cancels work without producing a final answer; saved progress and history remain available. Stopping the main task stops its entire child tree. Reopening unfinished work requires an explicit resume or stop. Back up the entire session directory: all agents share the main JSONL transcript, and task recovery also needs the adjacent state files.

Run locally

Install Vite+, then run:

vp install
vp run dev

To build a macOS installer:

vp run package:mac

The .dmg is saved to apps/desktop/dist/release/.

To build a Windows x64 installer on Windows:

vp run package:win

The .exe is saved to apps/desktop/dist/release/. Windows command execution requires Git for Windows with Git Bash installed.