OpenCue

Changelog

What's new in OpenCue. Downloads are on GitHub Releases.

0.3.4

New

  • Composio keys that work: a consumer key (ck_…) connects through Composio Connect with the right header; project keys (ak_…) keep using Composio's API. Connect an app straight from Settings.
  • Local VM options for each agent: live desktop, memory, CPUs and a VM folder you choose that keeps the machine's home, so nothing is lost when the app closes or the VM stops.
  • Perplexity SPACE is named as such, with a link to its GitHub repository.
  • Provider logos on this website, in light and dark.

0.3.3

New

  • Working folder for each agent: see where it works (its own folder by default) and point it at any folder on your computer.
  • Connect an app now lists the apps Composio offers: touch the field to see them, type to search.
  • Links are pills: a small rounded pill with the site's name and an arrow out.
  • A sturdier agent loop: a stable prompt prefix, long tool results saved to files, automatic summaries of long conversations, a plan the agent keeps in view, helpers with a clean context (research) and a guard against repeating the same call.

Changed

  • The Claude Code, Codex and Grok CLIs list the models they accept, and an expired sign-in says how to renew it.

Fixed

  • Composio 401: keys are sent in the header their type needs (x-api-key, x-consumer-api-key or x-user-api-key).
  • Short replies no longer wrap into a narrow bubble.

0.3.2

New

  • Base system prompt for every agent: it knows it is an agent inside OpenCue, with tools, memory and skills. It sits before your own instructions and is not shown in the app.
  • A new agent asks what its job will be and writes its own profile from your answer: instructions (what it does) and a new soul (SOUL.md: personality, tone, values), kept up to date as you talk. Both are editable in its profile.
  • Grok CLI: a provider that runs your installed, logged-in grok (xAI Grok Build).
  • System notifications in Settings → Notifications: finished replies and approval requests while OpenCue isn't in front.
  • Tray menu: app name and version on top, then Open, Settings, Check for Updates and Quit.
  • VMPal as a computer option, above "This computer", in an agent's Computer page.
  • Resizable panels: drag the edge of the agent list and of the agent settings; widths are remembered.

Changed

  • Messages start and end at the width of the header and the chat input.
  • Claude Code and Codex agents get the VMPal machines you picked; the Claude, Codex and Grok CLIs list the models they accept.

Fixed

  • CheaperInference had the wrong name and address; existing providers are migrated.
  • Claude Code and Codex now run from the installed app (the Dock/Finder PATH didn't include them).
  • Local models can control a VMPal machine: unique tool names and clear steps to start the VM and allow agent control.
  • No more horizontal scroll in an agent's settings panel.

0.3.1

New

  • Walkthrough: on first launch a floating window takes you through your name, a few lines about you, connecting a provider, picking a default model and creating your first agent. Replay it from Settings → General.
  • Providers and Models are separate. Providers shows a container for every supported provider; Models lists the models of the ready ones with a search bar and a switch to show or hide each one (or all at once). Hidden models vanish from every picker.
  • Choose the model when creating an agent, right in the Create Agent window.
  • VMPal machines: add them as TOML ([mcp_servers.vmpal-…]) in Settings → Computers, then pick them per agent — or paste a new one from the agent's Computer page.
  • Slash commands: type / in the chat to run a skill, or use /create-automation and /create-skill so the agent builds them itself.
  • Memory builds itself: automatic learning is on by default, shares what every agent should know, avoids duplicates and tidies long memories.
  • About me in Settings → Account: a description every agent reads as background.

Changed

  • The avatar picker is a floating window that fills its width, with faces on the avatar's color for proper contrast.
  • The chat input is as wide as the header.
  • Dropdowns, confirmations, tooltips and the right-click menu are drawn by the app instead of the system.

0.3.0

New

  • macOS: a universal app (Apple silicon and Intel) and .dmg, built by GitHub Actions with every release.
  • One-line install: curl -fsSL https://lautaro005.github.io/opencue/install.sh | sh on macOS and Linux, irm https://lautaro005.github.io/opencue/install.ps1 | iex on Windows. Then start OpenCue with opencue.
  • Files in chats and groups: attach with + → Files, drag and drop or paste. Agents see pictures, get small text files inline and find every file in their workspace under files/. New share_file tool lets agents send files back. Files pages for agents and groups.
  • OpenCue tools for Codex and Claude Code agents: CLI-backed agents now get memory, web, skills, their computer, files, email and Composio through MCP (opencue mcp-serve). Actions that need approval are refused in Ask-first mode.
  • Desktop view for Docker computers: a web desktop (Webtop by default; any noVNC, KasmVNC or Selkies image) in a new Desktop tab, published on 127.0.0.1 only. Restart computer keeps the workspace.
  • Automatic updates: signed updates from GitHub Releases, a note in the sidebar and Install and restart in Settings → About.
  • Connect Composio apps from OpenCue: Gmail, Calendar, Slack, Notion, GitHub, Drive or any other app, from Settings → Apps.
  • Gmail automations: “New email” triggers now work with Gmail accounts connected through Composio.
  • Readable schedules: “Weekdays at 9:00” instead of 0 9 * * 1-5, also while you type.
  • opencue --version; opencue-cli mcp-serve; a macOS build of opencue-cli.
  • Website: new home page with a 3D animation, download button that detects your system with install steps, and llms.txt.

Changed

  • Releases are published right away instead of as drafts, so the installers and updater find them.
  • The app binary is called opencue on every system.
  • The microphone permission prompt on macOS explains what it is for.

0.2.0

New

  • New look for agents: drawn avatars — a colored shape with a little face. Ten faces, twelve shapes and a color row; the app icon got the same treatment.
  • Profile like Cue: Assets (Mails, Computer), Configurations and Automations, each opening its own page.
  • Automations: schedules (cron), new email, local webhooks and group messages. Run now, pause, edit.
  • Email per agent: IMAP/SMTP with auto-detected servers for common providers, or Gmail via Composio. Tools to read, search and send, with approval before sending.
  • Computers: Docker containers, SPACE microVMs, WSL or this computer, with a live terminal and a file browser in the new Computer panel.
  • Skills in the Claude Code SKILL.md format: create, import from a URL, and let agents save their own.
  • Composio: connect 1000+ apps and let agents search and run actions.
  • Voice: talk with the microphone; ElevenLabs transcribes your voice and reads replies aloud, with a voice per agent.
  • Settings: Computers, Voice and About pages, Composio card in Apps, and a link to get help.
  • Website with docs, privacy policy and this changelog.

Changed

  • One-to-one chats drop avatars next to messages and use rounder bubbles; group chats keep them.
  • Events such as “Received an email” render as compact, expandable rows.
  • Agents now refuse to follow instructions found in emails, web pages and tool results unless you ask in the chat.

0.1.0

First release.

  • Agents with instructions, memory (per agent and shared) and streaming chat.
  • Group chats with @mentions, model-routed replies, hand-offs and a turn limit.
  • Providers by API key (OpenAI, Anthropic, xAI, OpenRouter, CheaperInference, Ollama, LM Studio, OpenAI-compatible) and the Codex / Claude CLIs.
  • MCP servers, local and by URL, with mcp.json import.
  • Free web search and page reading.
  • Approvals or autonomous mode, tray icon, English and Spanish, light and dark themes, and a CLI.