Skip to main content
API bodies are JSON (100 KB at most), so a file goes in steps: you ask for an upload, PUT the file to the URL you get, then complete it. CaptureFlow checks the file and creates the input, as an upload in the app does.
1

Create the upload

size_bytes is the exact size of the file. Optional: idea_id (one of your ideas), orientation (horizontal or vertical, for a video). The answer has an upload with url, method (PUT), headers and expires_at.
2

PUT the file

Send the whole file in one PUT to upload.url, with exactly upload.headers and no Authorization header:
Any 2xx means it arrived. The URL is private. For video and audio it only accepts this file’s type and size; for every kind, the size and type are checked again when you complete.
3

Complete it

CaptureFlow checks the size, the type and the file’s first bytes, then creates the input. The answer’s input has the input id. Completing again is safe: it answers with the same input.
4

Wait for processing (audio and video)

Call GET /v1/inputs/{id} until neither processing.media nor processing.transcript is pending. Each ends as ready, failed or none (nothing to expect: an image has no transcript, nor does a scanned PDF). Video and audio take a few minutes; PDFs and images are ready at once. Then draft from it, or start the content agent with its input_id.

​
Accepted files

HEIC and SVG images are not accepted. Text goes through POST /v1/inputs/text, web pages through POST /v1/inputs/url. The text of a PDF is read from its first 200 pages (200,000 characters at most); a scanned PDF has no text.

​
Timing and limits

  • Complete an upload within 24 hours. The upload URL lasts 6 hours (video, audio) or 2 hours (PDF, image); POST /v1/inputs/uploads/{id}/url gives a new one, or answers uploaded: true when the file is already there.
  • Unfinished uploads at once: 20 per key or connected app, 50 per workspace and 10 GB in total (an agency’s client workspaces share the agency’s).
  • Per day, per workspace (an agency’s client workspaces share the agency’s): 20 audio or video uploads, 100 PDFs or images, 20 GB. During a trial: 5, 25 and 2 GB. Over that: upload_budget_reached.
  • An upload belongs to the key or app that created it.

​
When something goes wrong

Retryable answers also carry details.retry_after_seconds.
Over MCP, the upload tools need an assistant that can run HTTP requests or a shell (Claude Code, Cursor, VS Code). Claude and ChatGPT in the browser can’t PUT a file, so these tools are hidden there.