file_tools is a Fetch Hive system MCP tool that is added to every agent automatically. It cannot be disabled or removed. Fetch Hive attaches the provider-facing file tools to a run only when the turn has document attachments, history file context, or asks to generate or export a file.
Uploaded CSV, XLSX, PDF, DOCX, text, and Markdown attachments are exposed to the agent as an <available_files> manifest. The agent should call read_file before relying on document contents. read_file can only access URLs that were supplied as structured attachments in the current request or conversation history.
Public API callers can pass attachments as HTTPS URL strings or as objects with
file_url, file_name, and file_type. URL strings are enough for most
calls; Fetch Hive infers metadata from the URL when it can.
When the current turn asks to generate or export a file, agents can call write_file to create PDF, DOCX, XLSX, or CSV files. For PDF and DOCX files, markdown content is recommended and preserves common document structure such as headings, paragraphs, lists, emphasis, code blocks, blockquotes, horizontal rules, and simple tables. For CSV files, provide CSV or JSON tabular content (input_format csv or json; JSON uses a single sheet). Generated files are stored as workspace assets and returned with a time-limited file_url (plus file_url_expires_at and asset id). Use the asset id with the assets API if you need a fresh download URL later. File tool calls are recorded in traces and completions as read_file and write_file sources.
When a Hive Agent run asks for a downloadable table or document, Fetch Hive produces the file after compose and attaches it to the run alongside the text final response (asset id, URL, name, and type). Retrying compose reuses already-verified node work without re-running the plan/execute/verify loop.
In a saved Agent chat, generated and uploaded files stay available in Use from this chat, even after older messages are summarized. Choosing one reuses the existing Asset without uploading it again. File contents are still read only when the Agent needs them.
