Before you start

What to have in place before installing Janus, and the settings in Claude Code and LM Studio that differ from their defaults. Each tool's own docs cover installing it; this page says what Janus needs from it.

What you need, by feature

You want	You need
The books (janus stats), the guard, captures	Claude Code, signed in
Subagents or other requests on a local model, a stand-in when Claude is down, worker tools, compression	Also LM Studio, a model that can drive Claude Code's tools, and the memory to run it
To build Janus yourself	Also Go 1.27 or later. Not needed for a release build

Start with the first row. Janus is useful on its own, and a week of janus stats tells you what there is to save before you set up a local model.

Claude Code

Install it and sign in as usual, with a subscription or an API key: Claude Code setup. Check that claude works on its own before putting Janus in front of it. Janus passes your login through as it is, so there is nothing to sign in to again, and nothing to change in your account.

Three settings point it at Janus, none of them a default:

Setting	Why
ANTHROPIC_BASE_URL=http://127.0.0.1:8787	Sends Claude Code's requests to Janus instead of to Anthropic.
ENABLE_TOOL_SEARCH=true	Claude Code turns tool search off when its base URL isn't Anthropic's, and then sends every tool's definition with every request. This turns it back on.
claude mcp add janus -- ...	Only for the worker tools, which Claude calls to hand reading to the local model.

Installing shows how to set the first two for one session, or for good.

LM Studio

Only for the local-model features. Janus talks to the LM Studio app (Bionic) over its HTTP server, using LM Studio's Anthropic-compatible /v1/messages and its model API. Install it from lmstudio.ai and see its docs for anything beyond what is here.

A model

Download one in the app. Janus is tested with qwen/qwen3.8-27b at 4-bit on a Mac with 36 GB of memory, where one request at a time is comfortable and two at once took it past 31 GB. Small models, around 4 billion parameters, could not drive Claude Code's tools in testing: their answers came back empty or rambling.

You don't have to load the model yourself. Janus asks LM Studio to load it when a request needs it, with a 65,536-token context unless you set contextTokens.

Settings that differ from the defaults

Setting	Where	Why
The server running	Start it from the app, and have it start when the app opens (autoStartOnLaunch in the server config file below), so it is there after a restart.	Janus reaches LM Studio only through its server, on port 1234 unless you change it.
Serving on the network	Only when LM Studio is on another machine. Turn on serving on the local network in the server settings, or, with the app quit, set networkInterface to 0.0.0.0 in ~/.lmstudio/apps/bionic/.internal/http-server-config.json.	Until then it listens on loopback only, so other machines can't reach it.
No API token required	The server settings, if your version can require one.	Janus sends no credentials to LM Studio.

Two things to avoid:

• Don't start the server with lms server start while using the Bionic app. The command-line tool tries to wake a separate headless service, which takes local inference away from the app and stops its server. Start the server from the app.
• Don't leave local.baseUrl at its default. http://studio.local:1234 is the Janus developer's own LM Studio machine, a placeholder. Set it to yours.

Where things run

One machine. Claude Code, Janus, and LM Studio all on your computer. local.baseUrl is http://127.0.0.1:1234, and nothing is exposed to the network.

LM Studio on another machine, such as a desktop with more memory. Claude Code and Janus stay on your computer; only LM Studio is on the other. On that machine:

1. Turn on serving on the network, as above.
2. If it has a firewall, allow incoming connections to LM Studio. With the macOS firewall on, macOS asks the first time the server listens on the network.
3. Find its name or address. On a Mac, the local hostname is in System Settings, General, Sharing, and ends in .local. A .local name depends on mDNS, which some routers and VPNs break; an IP address doesn't.

Then, from your computer, check that you can reach it before involving Janus:

curl http://mymac.local:1234/v1/models

It lists the models LM Studio has. Put the same address in local.baseUrl.

Janus itself always listens on loopback (127.0.0.1:8787), on the same machine as Claude Code. It forwards whatever credentials a client sends, so it refuses to listen on the network unless allowRemote says so.

Then

Install Janus, and run janus doctor. It checks the config, the port, that Anthropic is reachable, that LM Studio answers and has your model, and the log folder, and says what to fix.