{"version":"2.1.301","anchor":"mcp-tool-call-input-gains-result-not-for-model-for-programma","canonical_anchor":"mcp-tool-call-input-gains-result-not-for-model-for-programma","heading":"MCP results can skip the model, and the starting folder is withheld in restricted setups","tier":"use","area":"SDK","scope":"both","heads_up":false,"url":"https:\/\/changelogs.core-directive.com\/v\/2.1.301\/e\/mcp-tool-call-input-gains-result-not-for-model-for-programma","release_url":"https:\/\/changelogs.core-directive.com\/v\/2.1.301","markdown":"### MCP results can skip the model, and the starting folder is withheld in restricted setups\n\nHosts can get raw MCP results with result_not_for_model, and the original working directory is left out under HIPAA-style restrictions\n\n**What**\n\nAn MCP server is an add-on that gives Claude extra tools. The tool Claude Code uses to call them now accepts an optional `result_not_for_model` flag:\n\n- When set, the result comes back as the raw MCP content, with nothing resized, saved or shortened, and the `count_tokens` size check skipped.\n\n- A result longer than 10485760 characters fails instead.\n\n- It is honoured only for a host connected over the process's own pipes. Hosts can detect support through `mcp_call_result_not_for_model_v1`, which is advertised only when the connection neither redacts what it saves nor is remote.\n\nSeparately, the information Claude Code builds its prompt from gains an `originalCwd` field, the folder the session started in. It is left empty when a privacy check is true: when a `hipaa` restriction applies, when another restriction check is true, or when the check itself fails.\n\n**Why**\n\nPrograms that drive Claude Code can fetch large MCP results for their own use without them being cut down or sent to the model. In privacy-restricted setups the starting folder's path is kept out of the prompt.\n\n- Area: SDK\n- Names: `result_not_for_model`\n- Tier: Use it now\n- Useful: 4\/5\n- Signal: 3\/5\n- Scope: both\n- Heads-up: no"}