Cubebot

Feature tour

What runs today, in plain words. Everything on this page has run in automated tests or real installs; what has not is listed at the end.

Agents that last

  • One permanent conversation. No sessions to lose. Long histories are summarized automatically when the model's context fills up.
  • Always on. An agent keeps its sandbox running and sleeps while idle. Any message, routine or call wakes it.
  • A workspace that stays. Its files live on a persistent volume that survives restarts, upgrades and a replaced sandbox.
  • A journal before every action. Each tool call is written down before it runs. After a crash, unfinished calls are shown as uncertain instead of being run again blindly.
  • Steer or stop mid-turn. Send a correction while it works and it applies it at the next safe point. Cancelling also stops everything it delegated.
  • Parallel lanes. Besides its main conversation, an agent can work up to four delegated tasks at once.

A computer of its own

  • Terminal and files. A real Linux shell with Python, Node, git and build tools, plus reading, writing and editing files. Full output is kept for a week.
  • Browser and desktop. Chromium on a full desktop, driven through the page's accessibility tree or by native computer use with Anthropic and OpenAI models. Logins persist.
  • Take over at any time. Watch the desktop live, take the mouse and keyboard, hand it back, and the agent looks again before it continues.
  • Login hand-off. When a site needs you, the agent asks; you sign in or type the one-time code yourself.
  • A vault it cannot read. Passwords and secrets are typed into the page for the agent. The model never sees them.

Memory, dreams and skills

  • One memory per agent. Shared by every room and task. It saves what matters at the end of a turn, and you can search, edit or delete any memory in the console.
  • Dreams. While idle, an agent consolidates and corrects its memories and may write a new skill. Every dream leaves a report you can read.
  • A shared skill library. Skills are folders with a SKILL.md, revisioned and installed per agent. People, agents and dreams can all write them. A skill never grants permissions.
  • A knowledge base. One wiki for people and agents, with a block editor, Mermaid diagrams, version history and trash.

Working together

  • Rooms. Chat with several agents and people. Mention an agent or person by name, highlighted as you type, or let a coordinator pick who should answer. A room can start empty and gain members later.
  • Delegation. Agents hand work to each other and wait for the results, up to four levels deep.
  • Agents that build. Ask an agent to set up agents, rooms and teams. They can also ask to rename an agent or change its cube. Each request waits for a person to approve it on the turn that asked, and is then made as that person; the agent hears back only about requests that were declined or could not be made. A switch on each agent turns this off.
  • Teams. Agents grouped under a manager, with shared instructions, a title and role for each member, and a room of their own. Teams nest, an agent can be in several, and a task can go to a whole team.
  • Tasks. Describe what you want done. The orchestrator picks agents, writes each a brief and a checklist, and tracks the plan until it is finished or needs you.
  • Broadcast. One message to every agent, each answering in its own conversation.

Routines

  • On a schedule with cron in the routine's own time zone, on demand, from a webhook with its own secret, or on an account event like new mail or an assigned issue.
  • A routine is a prompt, a shell command, or both. A command alone runs without calling the model at all.
  • Agents can create and manage their own routines, up to 32 each.

Connections

  • 35 native connectors, from Gmail and GitHub to Home Assistant and Stripe. See the catalog.
  • Data: query PostgreSQL, MySQL and MariaDB, work with objects in S3, R2, B2 or MinIO, and files on any SFTP server.
  • Notifications: agents can buzz your phone through ntfy, Pushover or Gotify.
  • MCP servers, local ones running inside each sandbox and remote ones over Streamable HTTP.
  • A model per agent. Every agent follows the platform's default model unless you pick its own, from the provider in use or any provider with a saved key.
  • Six providers: OpenAI, Anthropic, Google Gemini, xAI Grok, OpenRouter and MiniMax, or a ChatGPT, SuperGrok or MiniMax M Plan.

Voice calls

  • Call an agent from the console. What you say joins its normal conversation and the answer is spoken back sentence by sentence. Interrupt it any time.
  • Speech runs on the server itself, offline after the first download (English), or through OpenAI or ElevenLabs. Pick a voice per agent.

The console

  • Live replies and every step. Each tool call appears as it happens: commands with their output, edits as diffs.
  • Files in chat. Pick, paste or drop up to ten images and files into a message. Agents see the images and can copy any file into their workspace.
  • Rich answers. HTML, SVG, Mermaid, tables and math render live in the chat, sandboxed from the network.
  • Home. Every agent at a glance, as a list or an org chart, beside what needs you, who is working and what it costs.
  • Calendar. Your calendars and every routine run in one month, week or day view. Create, move and delete events and routines from it.
  • Shared files. Upload, organize and share with people, groups or agents, with a viewer for images, PDF, audio, video, tables and code.
  • Activity inbox, command palette, light and dark themes, and an installable app that works on phones.

People and access

  • Password sign-in, invitations and single sign-on with OpenID Connect.
  • Groups, and per agent or room: no access, watch, chat or manage, set from the platform or from the agent itself.
  • A profile agents read: what to call you, your role, time zone, language and what they should know about you.
  • An audit log in plain sentences. Passwords, tokens and vault secrets never appear in it.

Usage and budgets

  • Tokens and estimated cost for every model call, per agent, per day and per model.
  • A daily and a monthly budget per agent. Past either, the agent pauses until the next day or month, or until you raise it.

Not yet

These are planned, and none of them is promised a date:

  • Hybrid memory retrieval
  • Skill imports from files and Git
  • Push notifications
  • Vendor webhooks for connectors

Being qualified now:

  • Real-cluster fencing and isolation
  • Network policy on production CNIs
  • Scale at 100 registered, 25 working agents
  • Live checks for every connector

Free, open and yours to keep.

Cubebot is Apache-2.0 and self-hosted, with no paid tier and no account to create. One person builds it in the open. A star helps others find it; a sponsorship buys the time to keep building.