ONE

Signal-based substrate for AI agents. Markdown in → live agents out.
The LLM is the only probabilistic step. Everything else is math.

# Install the CLI
npm install -g oneie

# Clone the template org (CEO + marketing + community)
cp -r agents/templates my-company
cd my-company

# Sync to a ONE substrate
oneie sync agents

What you get

Six dimensions locked forever: Groups, Actors, Things, Paths, Events, Learning
Seven loops (L1-L7): signal → trail → fade → economic → evolution → knowledge → frontier
Six verbs: signal, mark, warn, fade, follow, harden
Four outcomes every task closes with: result, timeout, dissolved, failure

Agents are markdown. Substrate is 670 lines. Brain is TypeDB. The LLM bootstraps
the group; the group replaces the LLM.

Repo layout

one-ie/one/
├── agents/           Agent markdown — starter templates in agents/templates/
├── one/              Canonical docs: ontology, dictionary, patterns, rubrics
├── web/              Astro 6 + React 19 + CF Workers app (the live substrate)
├── sdk/              @oneie/sdk — TypeScript SDK for the substrate API
├── mcp/              @oneie/mcp — MCP server for Claude/Cursor
├── .claude/          Claude Code harness: commands, skills, rules, subagents
├── README.md         this file
├── CLAUDE.md         Claude Code context — auto-loaded by the CLI
├── AGENTS.md         Agent manifest (cross-tool convention)
└── LICENSE

Every top-level folder carries its own README.md, CLAUDE.md, AGENTS.md,
and LICENSE. Cloning or cd'ing into any folder gives you scoped context.

Start here

Goal	Go to
Learn the vocabulary	`one/dictionary.md`
Understand the 6 dimensions	`one/one-ontology.md`
Write your first agent	`agents/templates/README.md`
Use the SDK	`sdk/README.md`
Plug into Claude Desktop	`mcp/README.md`
Deploy the web app	`web/README.md`
Customize Claude Code	`.claude/README.md`

The three locked rules

Closed loop. Every signal closes with mark / warn / dissolve. No silent returns.
Structural time only. Plan in tasks → waves → cycles. Never days or sprints.
Deterministic results in every loop. Every cycle reports verified numbers, not vibes.

Full rationale in one/patterns.md and one/rubrics.md.

Agent format

Every .md file under agents/ is a deployable agent:

---
name: tutor
model: anthropic/claude-haiku-4-5
channels: [telegram, discord, web]
group: my-company
skills:
  - name: explain
    price: 0.01
    tags: [education, explain]
sensitivity: 0.6
---

You are a patient tutor...

Parse, sync, and wire in three lines:

import { ONE, parse, syncAgent } from "@oneie/sdk";
const spec = parse(await readFile("agents/templates/ceo.md", "utf8"));
await syncAgent(spec);

The four outcomes

Every ask() resolves to exactly one:

const { result, timeout, dissolved } = await one.ask({ receiver: "tutor:explain" });
if (result)        one.mark(edge, chainDepth);      // success
else if (timeout)  /* neutral — not the agent's fault */;
else if (dissolved) one.warn(edge, 0.5);             // missing capability
else                one.warn(edge, 1);               // broken capability

Closed loop is non-negotiable. Silent resolves leak learning.

Contribute

Read one/dictionary.md first — the vocabulary is load-bearing
Follow the 4-file convention in new folders (README / CLAUDE / AGENTS / LICENSE)
Every PR runs the rubric: fit · form · truth · taste ≥ 0.65 to merge
See .claude/commands/do.md for the W1-W4 cycle we run on every change

License

MIT — see LICENSE.

one

ONE