A deployable, machine-enforced coordination layer for multi-agent AI teams. Ujima keeps everything that worked from two internal prototypes — folder-as-medium, agent sovereignty, snippet+reference, checkpoints — and moves everything mechanical into a deterministic, zero-dependency Python CLI so the protocol can't drift, race, or rot between sessions.
"LLMs make judgments; tools do bookkeeping. Any rule that can be enforced by code must never be enforced by prompt."
A surface is a whole place an agent — or Malkio — is in: a CLI session, a boot digest, a workspace on disk, a live fleet. These are rendered as full frames from the real CLI output, the real file layout, and one real deployed workspace, not mockups of a UI that doesn't exist.
The pieces Ujima is assembled from — grouped by the job they do, not by directory. 47 components across 5 categories: every CLI subcommand, every schema, every template, every Convene phase, and every design doc.
What this site is. Ujima has no UI — it is a protocol plus a 1,718-line zero-dependency CLI. There is nothing to screenshot. Every card below renders real CLI help text, real schema fields, real template excerpts and real doc copy, pulled from the files in this repo on 2026-07-18. The visual identity (dark panel chrome, the green/gold/red mark) is composed for this showcase — it is not extracted from a product design system, because the product does not have one. See README.md for the full faithful-vs-approximated accounting.