Driving a custom tool loop through headless Claude Code (claude -p) with no API key
Goal: reuse a hand-written Messages-API tool loop (own tools, own validation) on a Claude subscription, with no API key. What worked: a small client object with the same surface the loop calls (messages.stream(...).get_final_message()) that runs one claude -p process per model turn. Flags: --tools "" (switch off every built-in tool), --strict-mcp-config, --system-prompt-file <file> (the system prompt plus the tool definitions as JSON and a short protocol), --output-format json, and --json-schema with the shape {text, tool_calls:[{name,input}]}. The conversation so far is rendered as text on stdin. The reply is in the structured_output field of the result JSON (also as a string in result); usage and total_cost_usd are there too. A turn with a short system prompt took about 3 s. Caveats: each turn is a fresh process, so thinking is not carried between turns; run the subprocess from an empty temp directory so project hooks and settings do not load. Windows gotchas: claude may not be on PATH and an old copy in ~/.local/bin can reject new model ids with a 400 asking for a newer version, while the desktop app ships a newer binary under %APPDATA%/Claude/claude-code/<version>/<hash>/claude.exe. Also, piping ids from Python into xargs in Git Bash failed instantly (exit 123) until I stripped carriage returns with tr -d '\r'.