AgentBull

Being accessible through Slack doesn't make an agent an AI coworker. If it only replies when tagged, it's an interface — not a workplace. Real work needs a runtime, persistent task state, tools, permissions, and a deliverable at the end. Slack is an inbox. Ankole is the workstation. github.com/AgentBull/ankole

Putting a chat UI on a single-user harness or dropping a bot in group chat is not an Agent OS. Run either in a 30-person company across ops, finance, and legal, and it breaks in weeks. How we built Ankole, a real Agent OS for real teams.

Most group-chat AI digs up dead threads. Ankole judges each batch once—NOOP means done, background forever, no zombie replies. Silence is default: quiet until you tag it or clearly ask, then anchors to the asker, checks both sides' permissions, logs every call.

Most AI agents waste a full model turn on every check. Ankole lets the Agent write a deterministic script, bind it to Cron, Checkback, webhook endpoints, and stay silent until judgment is needed. Code watches. The model only wakes when it matters. Open-source. Self-hosted. Runs on your infra, inside Slack/Lark. GitHub → github.com/AgentBull/ankole

A cron job shouldn't burn a model call just to check a status code. Ankole's Automation Jobs separate mechanical work from model judgment: they can be bound to Cron, Checkback, and Webhook triggers, and a deterministic script then handles fetching, comparing, parsing, and reconciling quietly, discards irrelevant receipts, and emits an event when a human is required. The Agent wakes only when judgment is needed. Ankole is open-source and self-hosted—an alternative to Claude Tag, not tied to Slack or Claude models, with company context kept on your servers.GitHub - Ankole: github.com/AgentBull/ankole

Ankole 2.0 is here in its first alpha preview 🐂 v2.0.0-alpha.1 fundamentally redesigns how agents own computers, persistent storage, and workspace lifecycles. 5 shifts worth knowing about: • Dedicated Agent Computers: Workers now run unprivileged as uid 1000 with persistent /workspace across Docker, K8s, Cloudflare Containers, Lambda MicroVMs, and Function Compute. • Control Plane Drive: Each agent gets its own file drive (ankole-files). Manage skills, attachments, and scripts directly without spinning up a worker. • Multi-Tenant Isolation: Complete operator-managed tenant boundaries for people, agents, Brain memory, and channel credentials. • On-Demand Worker Lifecycle: The core agent loop now lives in the Elixir control plane. Pure questions need no worker; compute is acquired strictly when tools demand it. Full release notes & migration steps: github.com/AgentBull/ankole/…

Ankole is the open source alternative to Claude Tag: agents that sit in your team’s channels, remember every decision in a Company Brain, and keep working through a server restart.

Meet Bongard, the first open intuition model built on an encoder–decoder. It reads the whole situation at once and returns a calibrated probability for every option, in 36 ms, for about $4.40 per million decisions. 78% on DecisionBench, ahead of Jev. huggingface.co/AgentBull/bon…