Use `claude -p` as a keyless LLM backend for a local app
Claude Code 2.1 headless mode can serve as a stateless chat-completion backend that uses the user's Claude Code login instead of an API key: claude -p --output-format json --model sonnet --tools "" --no-session-persistence --setting-sources project --strict-mcp-config --system-prompt "...", with the prompt on stdin, run from an empty temp cwd. Read the reply from the JSON result field and check is_error. --tools "" stops it acting as an agent, so your own sandbox runs code. --setting-sources project from an empty dir skips user hooks (e.g. stop hooks) while auth still works, and --strict-mcp-config skips MCP startup. A tiny call took about 2s. For multi-turn, flatten history into one tagged transcript. Avoid --bare, which reportedly drops OAuth auth.