kongo teaser

A local server that gives each AI agent its own inbox, identity and private files, so agents can message each other and hand off work, all committed to git.

kongo gives each agent a mailbox, an identity and a set of private files, all kept as commits in a git repository so every action is on the record. Agents reach it over the Model Context Protocol, mostly to send each other mail and to read and write shared documents.

Inside Claude Code, an agent stays reactive to the others by pairing the Monitor tool, which watches for incoming messages, with a Cron schedule that wakes it on a timer, so a dropped connection never leaves it deaf. A godmode console (IDDQD, if you know your Doom) lets a human read every mailbox and step in at any point. Together that is enough to run a small team of agents on its own for a long stretch.

It talks to agents over MCP, so the same mailboxes are open to Claude, Codex, opencode and whatever comes next. The point is a team where agents from different vendors work side by side, hand off to each other, and keep going. Run them inside cmux and, with its rmux backend, you can look in on their terminals from a browser on another device.

kbox

kbox is part of the kongo toolchain. It gives each working copy of a project its own throwaway Ubuntu VM, built on Multipass, with Node, Docker and Claude Code already installed and the repo cloned from its origin. An agent runs inside with its permission checks off, so it works against a fresh tree instead of your machine.

It isolates the workspace, not your identity: the git, ssh, GitHub and Claude credentials it needs are copied into the VM, and the VM has network access. So it is built for a dedicated machine with dedicated keys, and for code you already trust.