Upload a binary asset (image, font, audio, …) to the project's hosted storage. This uploads bytes you actually hold — a file you generated, downloaded, or read yourself. Chat attachments don't qualify: the user's attachments never reach MCP servers (you see attached images through vision only; there is no file, id, or URL behind them you can read), so for those use request_user_upload instead and the user re-picks the file in a card that uploads from their browser. Three modes. ChatGPT conversation files — a generated image, a file ChatGPT itself holds: pass the file as the `file` parameter and the host attaches a download link itself; this server fetches the bytes directly, at full quality (nothing goes through your sandbox or through base64 in arguments; content_type and size_bytes are optional here). Never downscale or re-encode a generated image to fit the inline cap — pass it as `file` instead. Files up to 3 MB you hold yourself — pass content_base64 plus size_bytes (the decoded byte count) and the upload completes in this call, returning publicUrl. Larger files — pass size_bytes alone to get an uploadUrl; PUT the raw bytes to it with the same content_type and exact byte count (e.g. `curl -X PUT -H 'Content-Type: image/png' --data-binary @file.png '<uploadUrl>'`), then reference publicUrl. Some sandboxes (claude.ai Cowork, ChatGPT containers) block egress to S3: if the PUT fails in any way — connection failure, proxy error, or a response without an x-amz-request-id header — that block is permanent for the session, so switch paths instead of retrying or re-encoding smaller: the `file` parameter in ChatGPT for any file that exists in this conversation, content_base64 for files under 3 MB, request_user_upload for user-provided files, or a PUT from inside the project VM via run_code_in_vm (re-mint the URL first; it is short-lived). For AI imagery generated fresh, use generate_image. A single file can be at most 100 MB via the presigned mode (the inline content_base64 mode is capped at 3 MB).
