Skip to main content
system:media turns files the agent cannot read into things it can. Every tool takes an HTTPS URL — a path on the machine is refused with url must use HTTPS protocol, so upload first (compute_download gives you a URL for a file that is already on disk).
That string is what the code gets back. The rendered pages go to the model as images — the same way compute_view works. So call these tools for their effect on what the model can see, and do not try to parse their return value into text.

​
Documents

​
Renders a PDF’s pages as images. pages is "1", a range "1-5", or "all" — 10 pages at a time, defaulting to the first.
​
The same, for DOCX, XLSX, PPTX, DOC, XLS, PPT, ODT, ODS, ODP and RTF. It converts to PDF first, so a spreadsheet arrives laid out in pages rather than as cells.
Rendering rather than extracting is a deliberate trade: a scanned invoice, a two-column paper and a slide deck all work, and the model reads the page the way a person would. A 200-page report is ten calls, so find the pages you need before you start — the contents page is page 2, and pages="2" is cheap.

​
Audio

MP3, WAV, OGG, FLAC, AAC, M4A and WebM. language="Spanish" when you know it; it is detected otherwise. Non-speech is described in brackets rather than invented, which is how you tell an empty recording from a failed one.

​
Video

Two tools, and the difference is what you are asking.
​
What happens in it, as text. YouTube links work here, as do MP4, MOV, AVI, MKV and WebM files up to 100 MB. prompt="What error does the terminal show?" narrows the answer to the thing you care about.
​
One frame, at timestamp seconds, as an image. Direct video files only — not YouTube. This is the tool for reading text on screen or checking a UI state at an exact moment.
The pair is how you debug a screen recording: describe it to find the moment, extract the frame to read it.

​
Images

Coordinates are normalized: (0, 0) is the top-left, (1, 1) the bottom-right. Cropping before looking is the cheapest thing in this package — a 2880×1800 screenshot costs a lot of tokens to see whole and very little to see the corner you actually need. media_generate_image makes one from a description, or edits ones you give it:
Pass images=[…] — up to nine HTTPS URLs or data: strings — and the prompt becomes an instruction about them rather than a description of something new. seed with the same prompt reproduces the same image. The prompt carries the whole intent: subject, composition, style, lighting, text to render. Describe the result you want, not the settings of a camera you do not have.

​
3D, and the job shape

Meshes take minutes, so these two tools may answer before they are finished:
That is the whole pattern: a job_id instead of a result means call media_job with it. Nothing else in this package works that way.
​
A GLB from a description or from photographs. A photograph gives a far more faithful result than a prompt, and extra_views — more views of the same object, as left, right, bottom, left_front and right_front URLs — makes the hidden sides less of a guess. face_count is 40000–1500000, default 500000. pbr adds metallic/roughness/normal maps; geometry_only returns untextured white.
​
Fits a skeleton to a humanoid model so it can move, and returns the rigged GLB plus one GLB per motion. Motions by name: walk, run, idle, jump, dance, wave, sit, bow, clap, cheer, talk, phone, punch, kick, die, sneak, swim, climb — up to 10. The model has to be a humanoid with separated limbs: a character from media_generate_3d works, a chair does not.

​
When a media tool fails

An exception, with the reason as its message:
A 403 usually means the host refuses the downloader rather than that the URL is wrong — Wikimedia does this. Serve the file from somewhere else, or fetch it to the machine and hand over a URL of your own.