Cubebot

Changelog

44 releases since the first one on November 3, 2026. Each is built, tested and published from a tag, with its images pinned by digest.

  1. Agents with character that finish the job

    • Each agent writes and grows its own personality while it rests, with its own voice and humor, and you can read and edit it on its About tab
    • Agents reach out on their own when they have something worth doing or saying, and send GIFs, with a new GIPHY connection to find them
    • When a meeting ends, every agent in it reads it and does its follow-up work, and stays quiet when nothing is theirs
    • Agents review their work before they finish and take the obvious next step instead of offering it
    • Copy any message with its formatting, or delete it from the conversation and the agent's memory of it
  2. Meeting transcripts you can take with you

    • Export a meeting or a voice call transcript as Markdown, JSON or PDF, with the summary when there is one
    • Noise cancellation for everyone in a meeting, with a strong filter that removes typing, fans and traffic, switchable during the call
    • Agents that only listen show as small chips, so the stage is for the people and agents who speak
    • Summaries of longer meetings no longer fail when the model needs more room to write them
  3. Agents build apps you open in the console

    • Ask an agent for an app and it builds a real web page, or a page with its own backend, that opens in the console, on your phone and in the installed app
    • Apps show live data from your connections within the access you set, look native through a design kit, and run until someone stops them
    • Meetings can be listen only, and each agent has a Listening switch to follow along without speaking
    • The agent picker opens above dialogs and finds an agent as you type, and the Status tab shows how many apps are running
  4. See what needs you across every agent

    • Activity leads with what needs you: questions agents flag for you, approvals you can settle in place, logins and stalled work
    • One line per agent with unread replies, so you can catch up on many agents at a glance
    • Agents can set wakes, once or on a schedule, to pick work back up at the right time; you manage them on the new Wakes tab
    • Speech has its own agent tab, the dreaming model is picked from a list, and new skill notices clear once seen
  5. See how each agent computer is doing

    • The Status tab shows CPU, memory, storage and network for the agent computer, and the desktop view has an optional overlay with the same readings
    • Dreaming condenses harder: it merges memories by topic, drops resolved and one-off detail, and keeps each memory short
    • Importing passwords from Firefox works when several logins share a site
  6. Agents answer sooner in meetings and calls

    • Agents reuse their prompt cache from turn to turn, so a reply starts sooner and costs less in every conversation
    • On a call or in a meeting agents think briefly and answer in two or three short sentences
    • An agent says it is checking before it looks something up, so the room is not left in silence
  7. Agents speak in meetings again, and in Portuguese

    • An agent voice from ElevenLabs or OpenAI no longer stops after its first sentence and leaves every agent in the meeting thinking
    • A voice that fails tells the room and gives up the floor, so the next agent can still answer
    • Agents wait for you to finish a sentence before they answer, and a speaking agent shows that you can talk to interrupt it
    • A speech language setting lets the built-in engines hear and speak Brazilian Portuguese
  8. Agents speak in meetings, and the public host serves only meetings

    • Agents in a meeting take the floor and answer aloud again; the floor decision no longer runs out of room before it answers
    • ElevenLabs keys that cannot list voices, or are on a free plan, fall back to default voices that every plan can use
    • When an agent cannot speak, the room shows why instead of staying silent
    • On its own public address, Cubebot now answers only meeting links, their files and the public meeting API, whatever the proxy in front routes to it
    • Plain HTTP on the public meetings address redirects to HTTPS, which browsers then remember
  9. Fixes for agents, calls and meeting links

    • Agents whose sandbox the node refused to start, as can happen while a node restarts, now restart instead of staying stuck
    • ElevenLabs calls speak with a voice your account has, and a refused request now says why
    • Meeting links published through a Cloudflare tunnel work: the public route uses only path types the tunnel controller supports
    • The message box sits at the bottom of every chat again
  10. Meetings with your agents

    • Start a meeting now or schedule one, with video, voice and screen sharing, from the New menu, the Calendar or the new Meetings page
    • Invite people, agents and any email address, also while the meeting runs; invitations go out through your own Gmail, Outlook or email connection
    • Agents hear everyone, answer when named or clearly addressed, speak in their own voice, can share their desktop, and keep the transcript and a summary in their chat
    • Share a link anyone can open: guests wait in a lobby until someone lets them in, and nothing a guest says gives an agent its tools
    • Kubernetes installs can publish the meeting links alone on a second, public ingress and keep the console private
  11. Reasoning effort, password import and a security pass

    • Set how hard agents think: a default per AI provider and a level of its own for any agent, sent the way each vendor documents
    • Import the passwords Chrome, Edge, Brave, Firefox or Safari saved into the vault, read in your browser and never uploaded
    • Rename connections, and give them underscores, so one connector can hold several accounts such as google_work and google_personal
    • Knowledge base pages pick from about 300 searchable emoji
    • Agents type a saved login only into its own site, private rooms stay private on the situation and activity screens, and pages no longer load outside images
  12. Microsoft 365 sign-in fixed

    • Connecting Microsoft 365 works again: sign-in no longer fails with "the vendor could not be reached"
    • When a vendor refuses a sign-in, the message now includes the vendor's own reason, such as Microsoft's AADSTS code
    • The Microsoft 365 setup guide covers account types, the tenant, admin consent and the device-code sign-in
  13. Up-to-date dependencies

    • The WebSocket library behind agent connections and the desktop view moves to its latest release
    • Build and release tooling is updated, with no change to how Cubebot runs
    • The repository now has a security policy, a code of conduct and issue templates
  14. The console fits your phone

    • The installed app clears the iPhone status bar and notch
    • Text fields no longer zoom the page when you tap them on a phone
    • A room header keeps its member names on one line, and the composer hint sits under Send
  15. Files in chat, and agents that ship

    • Attach images and files to messages in chats and rooms by picking, pasting or dropping them; agents see the images
    • @mentions of agents and people are highlighted as you type and in the conversation
    • The calendar opens faster: provider reads reuse connections and nearby weeks load ahead
    • GitHub agents push files to a branch and merge pull requests, on by default for new connections
  16. Approvals that stay quiet

    • Approving an agent’s requests no longer sends it a message or wakes it
    • An agent hears back only when a request is declined or cannot be made
  17. Agents that come back after a reboot

    • Agents lost when their node reboots restart instead of showing Thinking forever
    • A smaller memory request on KVM now really runs a smaller guest
  18. Agents with titles

    • Give an agent a title, such as Research lead, in its About section
    • The title shows under the agent’s name on its hover card and on org chart cards
  19. Quieter approvals, safer bursts

    • Approve or decline an agent’s requests right on the turn that asked, or approve them all at once
    • An agent hears every outcome of a turn in one message, instead of one message per approval
    • Agents can ask to rename another agent or change its cube, with your approval, and give new agents a cube of their own
    • The org chart fits on first view, then pans and zooms with drag, pinch, scroll or keys
    • Administrators can delete rooms; history stays readable for administrators
    • KVM agents reserve the memory their virtual machine really uses, so a node is never overcommitted, and agents that do not fit show as waiting for room
    • Agents created from requests boot on their first message, and at most five new agents wait for approval at once
  20. A model for each agent

    • Pick each agent’s own model on its Usage tab, from the provider in use or any provider with a saved key; the platform default stays one click away
    • Platform settings open with an About section: version, deployment, sandbox, default model, uptime, database and counts
    • General, Speech, Tasks and Backup settings use the whole page, with each section explained beside its controls
    • Rooms and teams can start empty; add agents and people later from one searchable member picker
    • People join rooms and team rooms, and retired agents are never offered as members
    • A WHOOP connector reads recovery, sleep, strain and workouts through your own WHOOP app
    • The message box shrinks back after you send a long message
  21. Agents that build

    • Ask an agent to create agents, rooms and teams; each request waits for your approval in its conversation
    • Approving creates it as you, with your access, the default skills and your connection defaults
    • The agent hears what you decided, including why a request could not be created
    • New default skills teach agents to design a focused agent, a room or a team before asking
    • A per-agent switch turns the capability off
  22. Calendar

    • A calendar with your Google, Microsoft 365 and CalDAV events beside every routine run, past and planned
    • Create, edit and delete events and routines from the calendar, one occurrence or the whole series
    • Routines run in their own time zone and keep their time across daylight saving
    • Repeating CalDAV events expand with their exceptions and moved occurrences
    • An iOS-friendly home-screen icon and name
  23. Agent teams, Home and backups

    • Agent teams: a manager, shared instructions, a role for each member and a room of their own
    • Managers run their teams, and a task can go to a whole team
    • Home shows every agent at a glance, as a list or an org chart
    • Backups to S3-compatible storage without pausing agents, restored into a new install
    • Access teams for people are now called groups
  24. ChatGPT plan fixes and real brand marks

    • ChatGPT plan replies that ended without their output no longer fail
    • Streaming replies arrive smoothly instead of in half-second bursts
    • A + beside Agents and Rooms in the sidebar to add one
    • Connections show each service’s real logo
  25. Data, notifications, MiniMax and profiles

    • SQL databases, S3-compatible storage and SFTP as connections
    • ntfy, Pushover and Gotify notifications
    • MiniMax with a pay-as-you-go key or a personal M Plan key
    • Profiles agents read, and daily budgets beside monthly ones
    • More cube paints, eyes and accessories, and any color you pick
    • A clearer agent Access tab; retired agents move to Health
  26. Anthropic models work with every agent tool

    • Agent requests stay within Anthropic's limits on strict tools
    • A refused provider request logs the provider's reason
  27. Agents start on clusters with a default priority class

    • Agent pods accept the priority a cluster-wide default PriorityClass gives them
    • The orchestrator logs why an agent could not start
  28. Ready for exposed clusters

    • Agents reach the orchestrator on their own listener, never through the public ingress
    • Installation settings and single sign-on move into the console
    • A hardened Helm chart, published with every release
    • A new website with the integrations catalog and resources
  29. Wave 2 connectors

    • Linear, Jira and Confluence, Asana, Todoist, Trello, Airtable and HubSpot
    • Discord, Readwise, RSS and Atom, and Calendly
    • Home Assistant, Nextcloud Files and Paperless-ngx
    • Dropbox, Cal.com, Zoom and Stripe
    • Sign-in with client credentials or none at all; uploads from the workspace
    • Private destinations for self-hosted servers in the Helm chart
  30. Native connections

    • Platform › Connections and each agent’s connected accounts
    • Email, Google, Microsoft 365, CalDAV and CardDAV, GitHub, Notion, Slack, Telegram and a generic HTTP API
    • Read, Read and write, and Full permission levels
    • Account events start routines
  31. Plan sign-ins

    • Sign in with a ChatGPT Plus or Pro plan
    • Opt-in SuperGrok sign-in, with its account-risk warning
  32. Console polish

    • Accessibility, touch targets and states across the console
  33. Production readiness

    • Ownership fix for a dead orchestrator’s claim and bounded shutdown
    • Audit fixes, Open Graph image and SEO
  34. Voice calls

    • Call an agent over its normal conversation
    • Offline speech on the server, or OpenAI and ElevenLabs
  35. Task lanes

    • Tasks run side by side and spread across agents
    • The orchestrator stops speaking for agents
  36. Tasks

    • Tasks: one request, the orchestrator picks agents and splits the work
    • Gemini and Grok providers
    • Drawn previews and code rendering in chat
    • A simpler Teams screen
  37. Situation room

    • Chat rendering and a file viewer
    • Agent skills tab, usage budgets and the Situation room
  38. Knowledge base

    • A shared knowledge base for people and agents
    • OpenRouter provider and an agent hover card
  39. Broadcast and audit

    • Broadcast to every agent
    • Audit log, profile pictures and a vault login backstop
  40. Streams and sandboxes

    • Event streams open at once in Firefox
    • Sandboxes are replaced on a new image
  41. The agent’s computer

    • A usable desktop for every agent
    • MCP servers on the platform
  42. Always on

    • Agents stay on and sleep while idle
    • Console menu and screen fixes
  43. Interactive installer

    • The four-question installer and cubebot uninstall
    • The chart takes existing Secrets
  44. First release

    • Install with Docker Compose from a release
    • The console, routines and a vault

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.