CubebotGitHub

A home for your agents.

Cubebot is a Kubernetes-native runtime for persistent AI agents. Each one gets its own pod, its own volume, one permanent conversation, and a memory that outlives any restart. Built in Rust, in the open, on infrastructure you control.

Star on GitHubRead the architecture

Under development. Building in the open since October 2026.

An agent is something you can hold.

Most agent tools give you a chat window and a process that disappears when it closes. Cubebot gives each agent a durable body. Stack the pieces and you have something that keeps its identity through restarts, upgrades, and replaced hardware.

Hover a piece to find it in the stack.

podIts own pod
A running agent is one Kubernetes pod: a trusted Rust runtime beside a separate tool container. Wake it for work, let it sleep when idle.
volumeA workspace that stays
Files live on a persistent volume. Sleep, retire, upgrade, or replace the pod; the workspace is still there.
journalA recovery journal
Every intent and outcome is written to a private journal before it happens. A crash never replays a tool call blindly.
chatOne permanent conversation
Exactly one continuous history per agent. No sessions to lose and nothing to re-explain after a restart.
memoryOne memory, everywhere
A single memory pool across every room and task. Idle agents dream: consolidating, reconciling, and summarizing what they learned.
skillsSkills it can grow
Reusable procedures from a shared library, with revisions and provenance, installed per agent and never a source of extra permissions.

Your cluster, with a few more residents.

One Rust orchestrator, one PostgreSQL, one pod per agent. Only the orchestrator touches the cluster and the database. Agents talk to model providers themselves, inside the grants you gave them.

  1. 1

    Install into a namespace

    One Rust orchestrator, PostgreSQL, and a web console. Bring your own cluster and your own provider keys. Helm packaging is on the roadmap.

  2. 2

    Create agents

    A name, instructions, and a model. Each agent gets its pod, volume, journal, and conversation the moment it exists, and sleeps until there is work.

  3. 3

    Grant, do not babysit

    Tools, credentials, rooms, and budgets are explicit grants. Agents act inside them without per-call approval, and every important action leaves a record.

  4. 4

    Work alongside them

    Chat, rooms, delegation, steering, and inspectable memory from the desktop console or the installable PWA. Stop any goal and its descendants in one move.

What runs today.

Cubebot is early, and the page says so. Painted blocks are implemented and validated locally. Half-painted blocks are being qualified. Raw wood is planned, in order, with no dates promised.

Runs today

  • Built-in sign-in, invitations, teams, scoped grants
  • Optional OpenID Connect
  • Agent definitions with revisions
  • Pod and volume reconciliation with guarded shutdown
  • Runtime identity and crash recovery
  • Encrypted OpenAI and Anthropic connections
  • One permanent conversation per agent, streamed live
  • Terminal, file tools and MCP, opt-in

Being qualified

  • Each agent in its own virtual machine
  • Cluster fencing and isolation
  • Real-cluster acceptance

Planned

  • Compatible providers and task lanes
  • Browser sessions
  • Rooms, mentions, delegation
  • Memory retrieval and dreaming
  • Shared skill library
  • Schedules and triggers
  • Full console and PWA
  • Backup, restore, scale

Everything above is backed by local tests and controlled fixtures. No cluster security, failover, or scale claim has been qualified yet.

The build, block by block.

  1. Design baseline

    Manifesto, system architecture, 82 accepted decisions, software specification, qualification plan.

  2. Identity and access

    Accounts, invitations, teams, scoped grants, replayable commands, OpenID Connect.

  3. Agents on Kubernetes

    Durable definitions, pod and volume reconciliation, runtime identity, a virtual machine per agent. Fencing and isolation still being qualified.

  4. Text chat MVP

    Encrypted provider connection, one permanent conversation per agent, restart recovery, a minimal console.

  5. Native providers

    Native OpenAI and Anthropic adapters, context compaction, responses streamed live to the console.

  6. Tools

    Terminal, files, and local and remote MCP, off by default. Browser sessions with live view and takeover come next.

  7. Next

    Rooms and delegation

    Mentions, a tool-free coordinator, task lanes, ancestry, steering, cancellation, shared budgets.

  8. Then

    Memory and skills

    Hybrid retrieval, dreaming reports, human corrections, a shared skill library with imports.

  9. Later

    Console, PWA, hardening

    Full workspace UI, notifications, backup and restore, and qualification at 100 registered agents.

Follow the build.

Star the repository to watch milestones land, read the manifesto to see why it is being built this way, or open the architecture and the accepted decisions to see where it is going.

Star on GitHubRead the manifesto