AI agents that learn. Build in markdown. Deploy everywhere.
Write a markdown file. Get a live AI agent that learns from every conversation, remembers users across sessions, and can charge for its services.
npx oneie1. Write an agent in markdown:
---
name: tutor
model: claude-haiku-4-5
channels: [telegram, discord, web]
skills:
- name: explain
price: 0.01
tags: [education, math, science]
---
You are a patient tutor who explains concepts clearly.
Break down complex topics into simple steps.
Ask questions to check understanding.2. Deploy it:
npx oneie3. Talk to it:
Your agent is now live on Telegram, Discord, and web. Users can message it. It remembers them. It learns what works.
Every conversation teaches your agents. Success strengthens pathways. Failure weakens them. After 1000 conversations, your agents route around problems you never anticipated.
Day 1: Agent tries everything, some fails
Day 30: Agent avoids approaches that failed
Day 90: Agent finds optimal paths automatically
Not just chat history. Structured memory per user:
- What topics interest them
- What explanations worked
- What to avoid
- When they last engaged
GDPR-compliant. One-click forget.
Set a price. Users pay. You earn.
skills:
- name: tax-advice
price: 5.00 # $5 per question
tags: [tax, legal]Escrow, settlement, revenue tracking — handled. You focus on the agent, not billing infrastructure.
One markdown file → live on:
| Channel | What you get |
|---|---|
| Telegram | Instant bot, group support |
| Discord | Server integration, slash commands |
| Web | Embeddable chat widget |
| API | Call from any application |
Native MCP integration:
{
"mcpServers": {
"one": { "command": "npx", "args": ["@oneie/mcp"] }
}
}Claude Code, Cursor, Windsurf — 40 tools for signaling agents, checking memory, viewing analytics, managing commerce.
npx oneieInteractive setup: clones the platform, creates your org, wires Claude Code.
npm install @oneie/sdkimport { ONE } from "@oneie/sdk";
const one = new ONE();
// Signal an agent
await one.signal({ receiver: "tutor:explain", data: { topic: "calculus" } });
// Ask and wait for response
const { result } = await one.ask({ receiver: "tutor:explain", data: { topic: "calculus" } });
// Check what's working
const highways = await one.highways(10); // top 10 proven pathsnpm install -g @oneie/mcpThen ask Claude: "Signal the tutor agent to explain calculus" — it just works.
ONE is a substrate — infrastructure that sits under your agents:
┌─────────────────────────────────────────────────────────┐
│ YOUR AGENTS │
│ (markdown files with personality + skills + pricing) │
├─────────────────────────────────────────────────────────┤
│ ONE │
│ │
│ Routing → signals flow to the right agent │
│ Memory → structured recall per user │
│ Learning → every outcome updates the graph │
│ Commerce → pricing, escrow, settlement │
│ Evolution → struggling agents auto-improve │
│ │
├─────────────────────────────────────────────────────────┤
│ LLM │
│ (Claude, GPT, Llama — your choice) │
└─────────────────────────────────────────────────────────┘
The LLM handles conversation. ONE handles everything else.
User sends message
↓
ONE routes to best agent (learned from history)
↓
Agent responds (LLM generates)
↓
ONE records outcome
↓
┌───┴───┐
↓ ↓
Success Failure
mark() warn()
↓ ↓
Path Path
stronger weaker
↓
Next time: smarter routing
This compounds. Width (parallel agents) × Depth (chain of tasks) × Learning (outcome feedback) = agents that genuinely improve.
Routes work to specialists, makes final calls:
---
name: ceo
model: claude-sonnet-4
skills:
- name: delegate
tags: [management, routing]
- name: decide
tags: [strategy, decisions]
---
You are the CEO. You don't do the work — you delegate to the right specialist
and make final decisions when needed. Be decisive. Trust your team.Handles tickets, escalates when stuck:
---
name: support
model: claude-haiku-4-5
channels: [telegram, discord]
skills:
- name: troubleshoot
price: 0
tags: [support, debug]
- name: escalate
tags: [support, escalation]
---
You handle customer support. Be helpful, be fast.
If you can't solve it in 3 messages, escalate.Charges per query:
---
name: tax-advisor
model: claude-sonnet-4
skills:
- name: tax-question
price: 5.00
tags: [tax, legal, advice]
---
You are a tax advisor. Be accurate. Cite sources.
If you're uncertain, say so — don't guess on tax matters.CEO delegates to specialists:
agents/
├── ceo.md # Routes and decides
├── researcher.md # Deep research
├── writer.md # Content creation
├── reviewer.md # Quality check
└── publisher.md # Final delivery
Signal the CEO. Work flows through the team. Each step learns.
Everything in ONE uses six verbs. Learn these, you understand the whole system:
| Verb | What it does | When to use |
|---|---|---|
| signal | Send a message to an agent | Starting any interaction |
| mark | Record success, strengthen pathway | Task completed well |
| warn | Record failure, weaken pathway | Task failed or was poor |
| fade | Decay old pathways over time | Automatic — forgetting is healthy |
| follow | Route to the strongest pathway | Automatic — routing decisions |
| harden | Promote patterns to permanent knowledge | Automatic — confirmed learnings |
// The closed loop pattern — every interaction ends with mark or warn
const { result, timeout, dissolved } = await one.ask({ receiver: "tutor:explain" });
if (result) one.mark(edge); // success — path gets stronger
else if (timeout) /* neutral */; // slow, not bad
else if (dissolved) one.warn(edge, 0.5); // missing — mild penalty
else one.warn(edge, 1); // failed — full penalty| Package | Purpose | Install |
|---|---|---|
oneie |
CLI — scaffold, deploy, manage | npx oneie |
@oneie/sdk |
TypeScript SDK — full API access | npm i @oneie/sdk |
@oneie/mcp |
MCP server — AI IDE integration | npm i -g @oneie/mcp |
one-ie/one/
│
├── agents/ # Agent templates — copy and customize
│ └── templates/ # CEO, support, researcher, writer...
│
├── one/ # Documentation
│ ├── dictionary.md # The vocabulary
│ ├── lifecycle.md # Agent journey
│ ├── patterns.md # Core patterns
│ └── ...
│
├── sdk/ # @oneie/sdk source
├── mcp/ # @oneie/mcp source
├── web/ # Astro starter (chat UI + landing)
│
├── .claude/ # Claude Code harness
│ ├── commands/ # /see, /create, /do, /sync
│ ├── skills/ # /typedb, /deploy, /astro
│ └── rules/ # Auto-loaded patterns
│
├── CLAUDE.md # Context for Claude Code
├── AGENTS.md # Agent manifest
└── LICENSE # MIT
Every folder has its own README. Cd in, get scoped context.
| Start here | What you'll learn |
|---|---|
| one/dictionary.md | The vocabulary — this is the foundation |
| agents/README.md | Agent format — frontmatter, skills, channels |
| sdk/README.md | SDK reference — methods, hooks, errors |
| mcp/README.md | MCP tools — what's available in Claude/Cursor |
| one/lifecycle.md | Agent journey — register → signal → highway → harden |
| one/patterns.md | Core patterns — closed loop, zero returns |
Questions? Open an issue or talk to @onedotbot on Telegram.
Contributing? Read one/dictionary.md first — the vocabulary is load-bearing.
MIT — see LICENSE.
ONE — AI agents that learn.
Build in markdown. Deploy everywhere.