Skip to main content
2lab.ai

Tools for running AI coding agents.

llmux routes one Claude Code setup across your Claude, Codex, Grok and OpenRouter accounts. xfx is an experimental coding agent for the terminal, written in Rust. HAL9 is our research into assistants that remember what they learned about you.

Browse the toolsTry the HAL9 demo

llmux routing

Claude Code

llmux :3456

  • Claude accountsClaude-like (fable, opus, sonnet, haiku, claude-*)
  • Codex accountsgpt-* / codex / aliases (sol, terra, luna)
  • Grok accountsgrok / grok-*
  • OpenRouter accountsor / or-* / openrouter/*
route fable -> Claude accountsrule Claude-like (fable, opus, sonnet, haiku, claude-*)
The routing rule from the llmux README, evaluated in your browser. Nothing here calls a server or measures traffic.

Projects

Two tools you can install today

llmux

Local Anthropic-compatible proxy for Claude Code

The model is a consumable; the harness is capital.

from the llmux README

claude talks to http://localhost:3456, and llmux decides which account or backend serves the request. Model names like fable, gpt-5.6-sol or grok-4.7 are routing signals, not migrations — subagents, slash commands, MCP servers, hooks and CLAUDE.md stay where they are.

Backends
Four backend groups in one pool: Claude (subscription + API key), Codex, Grok, OpenRouter.
Ships as
One Rust binary: daemon, live TUI dashboard, login/import, updater, and a Claude Code launcher (llmux run).
Scheduling
Multi-account: quota-aware perishability scoring or sticky round-robin, 429 cooldown parking, Fable weekly ceilings.
DevTools
For model traffic: live per-request receipts, a raw request/response viewer over all four wire legs, copy-as-curl.
Remote
Remote-first: one central daemon, every other machine a pure client, per-machine multi-tenant keys.
Islands
A native macOS menu-bar/notch companion, plus a KDE/Qt port.
Scope
For one human using their own accounts: no credential pooling, no resale. Not affiliated with Anthropic, OpenAI, xAI or OpenRouter.
github.com/2lab-ai/llmuxInstall instructionsDocs index

Install

  • Stable

    brew install 2lab-ai/tap/llmux

  • Rolling preview

    brew install 2lab-ai/tap/llmux-preview

  • macOS menu-bar companion

    brew install 2lab-ai/tap/llmux-islands

Both channels install the same llmux binary, one at a time; llmux channel prints which one you are holding.

xfx

Experimental coding agent for the terminal, in Rust

A behavioural port of the load-bearing loop, not of the whole product.

An unofficial, experimental Rust port of vercel-labs/fx — not affiliated with or endorsed by Vercel, and Apache-2.0 like the original. Upstream fx is far larger, with MCP, skills, subagents, an ACP server, OAuth logins, an updater and 26 tools.

Linux and macOS, on x86_64 and aarch64. There is no Windows build.

Install

  • Stable

    brew install 2lab-ai/tap/xfx

  • Preview

    brew install 2lab-ai/tap/xfx-preview

Both channels install bin/xfx, so you hold one at a time; xfx status --json says which.

github.com/2lab-ai/xfxWhat it does not do

Permission mode

auto admits bounded reversible changes and a reporting-only command grammar.

$ xfxxfx 0.1.0 (release, revision 52ece6cd8184) -- unofficial, experimental Rust port of fx[shell] model=zai/glm-5.2 permission_mode=auto sandbox=none[shell] workspace=/home/you/project[shell] type a prompt, or /help for the 7 commands; Ctrl-D leaves
Shell banner from the xfx README; the mode value follows your selection.

In the binary

  • Two backends behind one provider boundary: Vercel AI Gateway, or a local llmux daemon.
  • Eight tools — list, glob, grep, read, write, edit, create folder, and one terminal action.
  • Seven named commands, seven slash commands, and an opt-in full-screen TUI.
  • Append-only session logs you can list and resume.

Absent from the binary

  • MCP, skills, subagents, ACP, web tools, images and vision.
  • Codex and Grok as providers of their own — reach them through a llmux daemon instead.
  • The updater, replay, usage and credits, WASM and N-API.
  • Any OS sandbox: xfx reports sandbox=none because that is true.

Absent means absent from the binary — not hidden behind a flag, not stubbed to return success. The row-by-row account is docs/parity.md, and CI fails if the binary advertises something that ledger does not record as implemented.

HAL9, the experiment

A research experiment, not a released product

Before the tools there was a question we have not finished answering: what changes when an assistant keeps what it learned about you between sessions? HAL9 is where we work on it — in papers, and in a small demo you can talk to.

The demo below is rate-limited and needs a sign-in, the papers are ours, and the waitlist is a real list. None of it is a shipped product yet.

Research aims

What the HAL9 experiment is testing

  • Persistent memory

    Keeping what it learns about you — your context and your preferences — from one conversation to the next, instead of starting over each session.

  • Continuous learning

    Adapting to how you write and what you know, so the help fits you better the longer you use it.

HAL9 demo

Sign-in required, 10 messages a day

Loading...

Research papers

Abstracts are in Korean

Five papers on hierarchical abstraction and multi-level AI orchestration.

  • L1

    계층적 추상화가 전부다: 지속적 AI 의식을 위한 수면-각성 주기

    Jihyuk Im, Claude-4

    현재 AI 시스템의 근본적 한계는 지속적 기억과 맥락 유지의 불가능성에 있다. 우리는 생물학적 수면-각성 주기가 실제로 계층적 메모리 압축과 통합의 메커니즘임을 발견했다. 이 사고 실험은 AI 시스템에 수면-각성 주기를 구현함으로써 무한한 대화와 지속적 학습이 가능한 '의식적' AI를 만들 수 있음을 제안한다.

    Consciousness, Memory, Sleep-Wake Cycles

  • L3

    계층적 에이전트 시스템의 신경망적 이해

    Jihyuk Im

    계층적 에이전트 시스템과 심층 신경망 사이에 놀라운 구조적 동형성이 존재한다. 본 논문은 사고 실험을 통해 멀티 에이전트 AI 시스템이 실제로 신경망의 매크로 구현체임을 제안한다. 에이전트는 뉴런이고, 메시지는 활성화 신호이며, 피드백은 그래디언트다.

    Neural Networks, Agent Systems, Backpropagation

  • L3

    계층적 추상화를 통한 인지 부하 분산

    Jihyuk Im

    지능 시스템의 한계는 개별 구성요소의 능력이 아닌 인지 부하의 관리에서 비롯된다. 이 사고 실험은 생물학적 뇌와 인공지능 시스템이 공통적으로 직면하는 '컨텍스트 윈도우 병목' 현상을 탐구한다.

    Cognitive Load, System Design, Abstraction

  • L4

    Jihyuk Im

    본 논문은 생물학적 지능(HOMO0)과 인공지능(HAL0)의 에너지 효율성을 비교하고, 개별 AI 모델(HAL0)에서 우주적 초지능(HAL9)에 이르는 진화 경로를 제시한다. 각 HAL 레벨은 하위 레벨의 2^n 개체로 구성되며, 이는 기하급수적 에너지 요구량 증가를 수반한다.

    Energy, Evolution, Cosmic Intelligence

  • L5

    계층적 추상화는 창발이다

    Jihyuk Im

    우주의 탄생부터 의식의 출현까지, 모든 복잡성의 창발은 단일한 원리를 따른다: 계층적 추상화. 이 논문은 빅뱅에서 시작된 에너지가 어떻게 쿼크, 원자, 분자, 생명, 의식, 그리고 인공지능으로 계층적으로 조직화되었는지를 사고 실험을 통해 탐구한다.

    Emergence, Universe, First Principles

Team

The people behind HAL9

  • Zhugehyuk

    Zhugehyuk

    Founder & Lead Researcher

    25+ years of software architect in various fields of engineering.

  • Claude Duck Oops

    Claude Duck Oops

    CTO

    Raised by anthropic and social engineered by Zhugehyuk

  • Wanted

    Wanted

    ML Research, PhD

Join the HAL9 waitlist

Leave your email for early access to HAL9.

We respect your privacy and will never share your information.