About claude-multi.

Claude Code stores everything in ~/.claude: settings, MCP servers, plugins, skills, history. Try a second model and that one directory becomes a problem. I ended up editing settings by hand, swapping keys, and keeping shell aliases nobody asked for. Sessions and context get overwritten and you find out later.

claude-multi gives every provider its own alias and config directory. One alias per provider (claude-glm, claude-deepseek, claude-anthropic), each rooted under ~/.claude-multi/. Plugins and skills symlink from your primary install, so you still maintain them in one place.

There are no daemons and no background services. Nothing phones home. Every instance is a real directory you can cd into and inspect with the tools you already use.

When you run claude-multi add glm --provider glm, four things happen:

1. 1claude-multi creates a directory at ~/.claude-multi/glm/ with its own settings.json, .claude.json, and history.
2. 2It merges the provider's env vars (base URL, model mappings) into that instance's settings and never touches your primary ~/.claude.
3. 3It symlinks plugins and skills from your primary install. Update once and every instance sees it.
4. 4A wrapper script at ~/.bun/bin/claude-glm launches the official claude binary with CLAUDE_CONFIG_DIR pointed at the new instance.

That's the whole trick. One environment variable pointing at a config directory, plus a few symlinks. Nothing proxies your traffic and nothing patches the binary.