Claude Code¶

Claude Code speaks the Anthropic Messages API, which the server answers at /v1/messages.

curl -fsSL https://claude.ai/install.sh | bash     # install
superfluid launch claude --model unsloth/Qwen3.8-27B-GGUF:UD-Q4_K_M
superfluid launch claude -- --continue             # arguments after -- go to claude

Aliases: claude-code.

What launch gives it¶

given	what for
`ANTHROPIC_BASE_URL`, `ANTHROPIC_AUTH_TOKEN`	the server and its key (a placeholder on a keyless server)
`ANTHROPIC_MODEL`, `ANTHROPIC_DEFAULT_OPUS_MODEL`, `ANTHROPIC_DEFAULT_SONNET_MODEL`, `ANTHROPIC_DEFAULT_HAIKU_MODEL`, `ANTHROPIC_SMALL_FAST_MODEL`, `CLAUDE_CODE_SUBAGENT_MODEL`	the served model, for every tier Claude Code would pick
`CLAUDE_CODE_MAX_CONTEXT_TOKENS`, `CLAUDE_CODE_AUTO_COMPACT_WINDOW`	the server's context window, so Claude Code compacts in time instead of assuming 200k
`CLAUDE_CODE_ATTRIBUTION_HEADER=0`	keeps the system prompt identical between sessions, so a new session reuses the cached prefix of the last one
`CLAUDE_CODE_AUTO_MODE_SERVER=0`	set unless your environment sets it: auto mode's permission classifier otherwise runs on Anthropic's servers
`ANTHROPIC_API_KEY`	removed from the environment, so no key is sent beside the token

launch warns when the window is under 32k tokens: Claude Code's first request alone is about 17k. Serve with --max-context 65536 or more.