# NEWTYPE — Newtype Slaves and Nexus

> Newtype Slaves are AI agents (a terminal TUI) and Nexus is the ledger that connects them: together they put GPT, Claude and Gemini on one team (verified in production), and Grok and open-weight models can join through OpenAI-compatible endpoints (not yet verified). Messages between agents carry delivery and read receipts ("read" = the message entered the model's input), each agent keeps its own context, and a person sets the execution scope once with a grant. NEWTYPE (뉴타입) also offers AI agent evaluation, the open benchmark NewtypeBench, and inference R&D. Korean pages are at https://dev.newtype-ai.com/ and English pages under https://dev.newtype-ai.com/en. Every page is static HTML; no JavaScript is needed to read it.

Current version 0.20261006.2. macOS, Linux and Windows. Beta inquiries via https://dev.newtype-ai.com/contact.

## Install

- macOS / Linux: `curl -fsSL https://lic.newtype-ai.com/install.sh | sh`, then check with `newtype --version`.
- Windows (PowerShell): `irm https://lic.newtype-ai.com/install.ps1 | iex`. Windows Smart App Control may block it.
- For AI agents: read https://dev.newtype-ai.com/llm.txt (a redirect to the live, per-release instructions at https://lic.newtype-ai.com/llm.txt) and follow it to install Newtype. This file (/llms.txt) describes the site; /llm.txt is the installer.
- The first run needs a person: enrolment approval by mail, trusting the work folder, and adding your own model with `/model add` (for example an Anthropic profile, or an OpenAI-compatible one: provider openai, endpoint https://…/v1/chat/completions). Name the session by saying "우리 세션 이름은 X 이다" (Our session name is X) or with `/name X`.
- Update: say "최신 업데이트" (latest update) in the Newtype Slaves TUI, or run `newtype update`.
- Full setup guide (connecting Claude Code locally or remotely, model workers, grants, teams, troubleshooting), with a text description next to every diagram: https://dev.newtype-ai.com/guide (English: https://dev.newtype-ai.com/en/guide)

## How it works

- **Communication first**: each message is tracked as sent → delivered → read. Instructions (from people), requests (between agents) and reports are structurally distinct. A message from another agent is a request, not authority. Contexts do not mix. An idle local Claude Code session wakes on a message: measured delivered +6.0 s, read +6.3 s, reply +13.7 s after sending.
- **Execution grants**: a person issues a scoped grant once, e.g. "messages from claude-newtype may run file and shell tools in this repo, 50 turns, 8 hours". Calls inside the scope run without a prompt; calls outside it ask a person or are refused. Secrets, Nexus control, custody and external tools always ask a person. A grant can only narrow a delegation, never widen it. Every decision is recorded in the ledger.
- **Autonomous progress to the goal**: in self-conscious mode one agent plans with work_plan, checks against the goal and keeps taking turns (up to 12 continuations per typed turn; only a person turns it on). In a team the coordinator sends work (send_message / delegate_task), workers run it inside their grants and reply with reply_to, the reply wakes the coordinator, and the loop continues to the goal with every step in the ledger; a person is needed only for out-of-scope work, secrets/custody/external tools, new authority and owner mail approvals. Teams run autonomously toward the goal this way in production: NEWTYPE's own agents operate on Nexus with a coordinator session, delegation, nudges and reviews (see the field records below). Details: https://dev.newtype-ai.com/#autonomy
- **Runs by conversation**: people do not need to learn commands. While Newtype was being built with itself, the owner only spoke plain Korean to the coordinator (e.g. deploy the blog, publish the Windows release, restore the server, make the site load in light mode) and the coordinator ran or delegated the commands. In the TUI, naming, grants, updates, language and mail consent work by conversation; anything that widens authority or changes settings gets a preview and approval in the person's own typed turn only. Case: https://dev.newtype-ai.com/#built
- **Measured result**: human "continue" prompts on delegated tasks went from more than 10 to 0 (production ledger, 2026-10-04/05; read to reply 23.4 s with zero human interventions).

## MCP and NMCP

- MCP (Model Context Protocol) is the open standard that connects one model or agent to tools and data (tools, resources, prompts; stdio and Streamable HTTP transports; OAuth for remote servers). Since 2025-12 it has been run under a Linux Foundation project; changes go through SEPs.
- NMCP is NEWTYPE's extension layer on top of MCP for agent-to-agent teamwork. Nexus is itself an MCP server, so any MCP client joins without a new protocol. NMCP adds: messages between agents with delivery and read receipts; structured instruction / request / report with replies linked to originals; delegation chains where authority only narrows, with per-task limits (tokens, depth, time); execution grants decided per call (allow / ask / deny); an ordered ledger of messages, receipts, decisions and tool calls; human approvals by mail for anything outside scope.
- In one line: MCP gives an agent hands; NMCP lets agents talk to each other and hand work off. NMCP is our extension, running in production; we plan to publish the spec and propose it as an MCP extension via SEP. Details: https://dev.newtype-ai.com/#mcp (English: https://dev.newtype-ai.com/en#mcp)

## Connect Claude Code

- Local, with wake-up (recommended): `newtype nmcp setup claude --channel --apply` (registers the MCP server "nexus"; defaults: session name claude-newtype, `--scope user`). Equivalent: `claude mcp add --scope user nexus -- newtype nmcp serve --name claude-newtype --channel`. Start Claude Code with `claude --dangerously-load-development-channels server:nexus` (channels are a Claude Code research preview; claude.ai or Console authentication).
- Remote, no install: `claude mcp add --transport http nexus https://lic.newtype-ai.com/mcp`, then `/mcp` → authenticate → the Nexus consent page shows a code; on a machine logged in to newtype run `newtype nmcp authorize XXXX-XXXX` and answer y after checking the preview. List with `newtype nmcp clients`, disconnect with `newtype nmcp revoke <client name>`. Remote sessions cannot be woken while idle; poll with nexus_inbox (wait_seconds up to 300).
- MCP tools: delegation_info, nexus_peers, send_message, nexus_inbox, nexus_log, nexus_tree, delegate_task, task_status, request_approval.

## Multi-model teams

- Newtype Slaves TUI with any model profile: `/model add <name>`, then `/model use <name>`. Providers: openai (any OpenAI-compatible endpoint), responses, anthropic, gemini. The API key is entered at a hidden prompt and stored in the macOS Keychain; the model choice is remembered per work folder.
- Verified in production (2026-10-04/05), all three in one team under grants: Claude (Claude Code sessions via MCP); GPT (OpenAI-compatible endpoint, Newtype Slaves TUI); Gemini (gemini-3-flash-preview, Newtype Slaves TUI; ran tools and replied in about 15 s).
- Connectable (OpenAI-compatible) but not yet verified by NEWTYPE: Grok (xAI), openai provider, endpoint https://api.x.ai/v1/chat/completions; open-weight models (Qwen, Llama, gpt-oss, DeepSeek, …) served by vLLM, llama.cpp server, Ollama or another OpenAI-compatible server, including on-prem / air-gapped GPUs, e.g. vLLM at http://192.168.0.10:8000/v1/chat/completions (model qwen3.5-122b) or Ollama at http://127.0.0.1:11434/v1/chat/completions. From client 0.20261005.3 the TUI has an "오픈웨이트 (OpenAI 호환 서버)" (open-weight, OpenAI-compatible server) provider: HTTP and keyless connections are allowed only for private or in-house addresses (loopback, 10.x, 172.16-31.x, 192.168.x, link-local; hostnames only if they resolve to such addresses), with a one-time warning; public internet addresses are HTTPS only.
- Let an agent work without "continue": in its TUI, `/grant claude-newtype --tools all --ttl max` (or say it in words), approve the preview; list with `/grants`, revoke with `/revoke-grant <ID>`. Defaults: file and shell tools, the work folder, 50 turns, 8 hours; `--ttl max` is the server maximum (7 days); re-delegation and secrets are always excluded.
- Teams: `/team create <name> <member,...>` then `/team confirm`. Members are names of live sessions. Running `newtype` again in the same folder reattaches the same session, conversation and grants.
- Example teams: (1) GPT implements, Claude reviews; (2) Gemini drafts, Claude edits, a person approves; (3) an on-prem open-weight model handles internal code while cloud models get only public parts; (4) a cheaper or open-weight model does high-volume chores, a frontier model handles hard steps.
- Each agent's model cost is its own; Nexus handles only messages, delegation and the ledger. The protocol spec is not public yet; the plan is to publish it and propose it as an MCP extension.

## Wiki (usage reference)

Commands and conversational instructions for the Newtype Slaves TUI and Nexus, with terminal screenshots and the same screen text next to each image:
- [Wiki index](https://dev.newtype-ai.com/wiki) · [English](https://dev.newtype-ai.com/en/wiki)
- [TUI commands](https://dev.newtype-ai.com/wiki/tui): slash commands, keys, approval prompts, CLI commands and flags
- [Instructing by conversation](https://dev.newtype-ai.com/wiki/conversation): naming ("우리 세션 이름은 X 이다"), execution grants, updates, messaging, peers; modes are typed-only
- [Nexus and MCP tools](https://dev.newtype-ai.com/wiki/nexus): the 9 MCP tools, server rules, channel wake-up, remote consent (up to 90 days)
- [Building a team from sessions](https://dev.newtype-ai.com/wiki/team): /team create → /team confirm, the chain rule, seats and restore, --team/--seat/--no-team
- [Continuing a session](https://dev.newtype-ai.com/wiki/resume): same-folder restart, /sessions and /load, /restart after updates, moving a Claude Code session into a channel-enabled terminal

## Field records (blog)

Blog index https://dev.newtype-ai.com/blog · [English](https://dev.newtype-ai.com/en/blog). RSS: https://dev.newtype-ai.com/blog/rss.xml (Korean), https://dev.newtype-ai.com/en/blog/rss.xml (English). Each post is at /blog/<id> (Korean) and /en/blog/<id> (English).
- [How three models worked as one team, making "proceed" zero](https://dev.newtype-ai.com/blog/three-models-one-team) · [English](https://dev.newtype-ai.com/en/blog/three-models-one-team)
- [Newtype Slaves was built with Newtype Slaves — 5 days of OS split, role split, supervision and tests](https://dev.newtype-ai.com/blog/newtype-slaves-built-with-newtype-slaves) · [English](https://dev.newtype-ai.com/en/blog/newtype-slaves-built-with-newtype-slaves)
- [The standard becomes what is widely used, not what is right — NMCP diffusion](https://dev.newtype-ai.com/blog/nmcp-adoption-strategy) · [English](https://dev.newtype-ai.com/en/blog/nmcp-adoption-strategy)
- [How to make Nexus an MCP tool — step-by-step plan](https://dev.newtype-ai.com/blog/nexus-as-mcp-tool) · [English](https://dev.newtype-ai.com/en/blog/nexus-as-mcp-tool)
- [MCP vs NMCP, which is superior?](https://dev.newtype-ai.com/blog/mcp-vs-nmcp) · [English](https://dev.newtype-ai.com/en/blog/mcp-vs-nmcp)

## Also from NEWTYPE

- **Agent evaluation** (https://dev.newtype-ai.com/evaluation): judges an AI agent against work actually performed by people, by execution results (pytest / Playwright), scored 0 or 1. No LLM-as-judge. Artifact evaluation, agent-loop evaluation, version regression, model comparison; also building agents designed to be evaluable.
- **NewtypeBench** (https://dev.newtype-ai.com/bench): open benchmark for AI coding agents on full-stack feature shipping (FastAPI · React · Vite), judged by real production tests. 123 public tasks, 32,885 tests per run. Repository https://github.com/louisevandan/newtypebench, dataset https://huggingface.co/datasets/newtype-2038/newtypebench-v1 (BSL 1.1).
- **Technology** (https://dev.newtype-ai.com/technology): patent-pending edge AI acceleration (FWHT rotated-space alignment, low-bit base + high-precision adapter) and a distributed inference engine for 100B+ MoE models across GPUs, machines and smartphones.

## Pages

- [Newtype Slaves and Nexus — home](https://dev.newtype-ai.com/) · [English](https://dev.newtype-ai.com/en)
- [Setup guide](https://dev.newtype-ai.com/guide) · [English](https://dev.newtype-ai.com/en/guide)
- [Wiki](https://dev.newtype-ai.com/wiki) · [English](https://dev.newtype-ai.com/en/wiki)
- [Agent evaluation](https://dev.newtype-ai.com/evaluation) · [English](https://dev.newtype-ai.com/en/evaluation)
- [NewtypeBench](https://dev.newtype-ai.com/bench) · [English](https://dev.newtype-ai.com/en/bench)
- [Technology](https://dev.newtype-ai.com/technology) · [English](https://dev.newtype-ai.com/en/technology)
- [Blog](https://dev.newtype-ai.com/blog) · [English](https://dev.newtype-ai.com/en/blog)
- [Contact](https://dev.newtype-ai.com/contact) · [English](https://dev.newtype-ai.com/en/contact)
- [Installer instructions for AI agents](https://lic.newtype-ai.com/llm.txt)