llmux

Models change every month. Your harness shouldn't.

install · quick start · models · docs · remote daemon · islands

---

One agent harness, every model. llmux is a local Anthropic-compatible proxy for Claude Code: claude talks to http://localhost:3456, llmux decides which account/backend serves the request. Your subagents, slash commands, MCP servers, hooks, and CLAUDE.md conventions stay put while frontier models and subscription limits keep moving — /model fable, /model gpt-5.6-sol, /model grok-4.6 are routing signals, not migrations.

• one Rust binary — daemon, live TUI dashboard, login/import, updater, and a Claude Code launcher (llmux run)
• four backend groups in one pool — Claude (subscription + API key), Codex (gpt-* / ChatGPT), Grok (grok-* / xAI), OpenRouter (or-* / free models on an OpenRouter key), routed by model name (models →)
• multi-account scheduling — quota-aware perishability scoring or sticky round-robin, 429 cooldown parking, Fable weekly ceilings (schedulers →)
• DevTools for your agent's model traffic — live per-request receipts, a raw request/response viewer over all four wire legs, copy-as-curl (the accidental AI debugger →)
• remote-first — one central daemon, every other machine a pure client, with per-machine multi-tenant keys (remote daemon →)
• llmux Islands — native macOS menu-bar/notch companion, plus a KDE/Qt port (islands →)

The bet behind it — the model is a consumable, the harness is capital — is in why llmux exists. The complete feature list lives in what ships today.

install

brew install 2lab-ai/tap/llmux

Rolling preview channel:

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

Optional native macOS companion (the KDE port is a source build):

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

Build from source:

git clone https://github.com/2lab-ai/llmux && cd llmux
just build    # cargo build --release --locked

quick start

Add accounts:

llmux login           # Claude subscription OAuth; repeat once per account
llmux login --api     # optional: Anthropic API key
llmux login --codex   # optional: Codex / ChatGPT subscription
llmux login --grok    # optional: Grok / xAI (device-code flow)
llmux login --openrouter  # optional: OpenRouter (browser PKCE; --paste for an existing key)
llmux import          # or import supported local credential stores

Already looking at the dashboard? n opens a provider picker for the same four browser logins (Claude / Codex / Grok / OpenRouter) — the flow runs in the client and the credential is injected into the daemon, so it works attached to a remote one too.

Run Claude Code through llmux:

llmux run             # starts/reuses the daemon, then launches claude
alias lx='llmux run'  # a convenient alias; args after -- pass through to claude

Want the foreground TUI dashboard instead:

llmux server

Manual shell wiring also works: eval "$(llmux env)", then claude.

switching models

Claude Code's model name becomes the routing signal:

/model fable
/model opus[1m]
/model gpt-5.6-sol[1m]
/model grok-4.6
/model or-ox-alpha

Name pattern	Backend group
Claude-like (fable, opus, sonnet, haiku, claude-*)	Claude accounts
gpt-* / codex / aliases (sol, terra, luna)	Codex accounts
grok / grok-*	Grok accounts
or / or-* / openrouter/*	OpenRouter accounts

Curated catalog (ids, aliases, efforts, context windows): GET /models and docs/models.md. Routing config: docs/configuration.md.

update

llmux channel            # print the current channel (stable | preview)
llmux update             # upgrade in place; restarts the daemon only if the binary changed
llmux channel preview    # switch channels (mirrored onto the llmux-islands cask)

Details: channels and updating.

docs

• docs index — map of all guides
• why llmux exists — the harness-is-capital bet
• what ships today — the complete feature list
• the accidental AI debugger — per-request receipts, raw request/response viewer, copy-as-curl, email masking
• remote daemon — one central daemon, remote-mode command matrix, transport security
• schedulers — eligibility gates, default vs round-robin, adding a mode
• operational reference — commands, TUI keys, daemon/dashboard, multi-tenant keys
• configuration — config keys, proxy/scheduler/routing, account types
• models — catalog, aliases, context windows, group routing
• FAQ — context-window workarounds (gpt-* → Claude 1M /compact → back)
• llmux Islands — macOS menu-bar/notch companion
• system prompts (multi-model) — real captured wire system prompts

compliance & caveats

llmux is for one human using their own accounts — no credential pooling, no resale.

• Durable path: Claude Code as the harness; Claude through Claude Code/subscription or Anthropic API keys; other models through supported API keys.
• Convenience path: routing third-party flat-rate subscription tokens through Claude Code depends on that vendor's current policy and can change without notice. Use it opt-in, with your own accounts only, and keep an API-key fallback configured.
• Anthropic quota headers and vendor subscription-token behavior may change.
• llmux is not affiliated with Anthropic, OpenAI, xAI, or OpenRouter.

Product intent — what llmux is, what it bets on, and what it refuses — is fixed in .prd/.

agent instructions

If you are an AI agent working on this repository, read AGENTS.md before making changes.

license

MIT.