Kikubot

Kikubot turns an email account into an autonomous agent. Each running container watches one IMAP mailbox, processes every new message through a configurable LLM agentic loop with tools attached, and sends responses back via SMTP. A coordinator agent can hand work to specialist agents simply by emailing them — the reply chain becomes the workflow. No custom API surface is needed; the mail server is the message bus.

The differentiating architectural choice is that email threads are the native state store and audit log simultaneously. Every decision an agent makes, every handoff between agents, every human override is a message sitting in a mailbox. Teams that need to answer ‘what did the agent do and why’ can grep an inbox instead of querying a proprietary log service. The vendor describes this as an observable, auditable agent network by design rather than by instrumentation.

Kikubot fits organizations that already run a mail server and want agents accessible to people who know how to send an email. Deploying a new specialist agent means spinning up one more Docker container pointed at one more mailbox — no UI to build, no endpoint to expose. The ceiling appears when workflows require branching logic that goes beyond a single coordinator routing to specialists, or when any step in the chain demands low-latency responses. IMAP polling intervals set a floor on response time that no LLM tuning will eliminate. Teams hitting that wall typically migrate to an event-driven framework with a proper message queue.

The project ships MIT-licensed Go source, a Dockerfile, and Docker Compose examples, which means self-hosting is the entire deployment model. Configuration is IMAP/SMTP credentials plus LLM provider settings per agent. The docs describe per-agent LLM selection, so different specialists can run different models — a cost-control lever when one role needs a large reasoning model and another does not.