Gemini web media analysis

Attaches a local image or video to the Gemini web composer and reads the reply. Verified working on gemini.google.com/app, Chrome, signed-in account, 2026-08-15.

Needs the claude-in-chrome MCP server — navigate, find, file_upload, computer, javascript_tool, get_page_text — and a Chrome profile already signed in to Gemini. No API key.

Two different paths. Use the right one:

Never click "Dateien hochladen" / "Upload files" in the uploads menu. It opens the native OS file dialog, which blocks the whole automation session — no further tool call reaches the browser until a human dismisses it by hand.

Video path

1. Navigate to https://gemini.google.com/app, wait for load.
2. Real click the + / "Uploads & Tools" button in the composer. A JS .click() does not open the menu — it needs real pointer events. The hidden input only exists while this menu is open.
3. find("hidden file input element") → returns Gemini's own input ref. Make sure no helper input of yours is in the page — it shadows this search and find returns yours instead. Remove it first. The uploader component does not mount on every menu open. If find comes back empty, close and re-open the menu and retry.
4. file_upload with the absolute video path and that ref.
5. Real click into the composer text area. This is the trigger — Angular picks the file up on the next tick.
6. Wait, and be patient. The thumbnail appears later than you expect — well beyond 8 s for a small clip. It renders as a still frame with a duration badge like ▶ 0:06. Do not conclude failure early. That mistake cost a whole session once: the upload had in fact worked and was reported as impossible.
7. Type the prompt and send, then poll get_page_text for the reply. Video analysis takes 30–60 s.

Image path

Gemini's own input rejects images — its accept list is documents and code only — so images go in through a synthetic paste instead. Build your own file input in the page, let the browser tool fill it with the real file, then hand that File object to the composer as paste data. Paste is the one mechanism Gemini's composer honours for images.

1. Navigate to https://gemini.google.com/app and wait for load.
2. Inject a helper input via the JavaScript tool:

let el = document.getElementById('__cin_up');
if (!el) {
  el = document.createElement('input');
  el.type = 'file';
  el.id = '__cin_up';
  el.setAttribute('accept', '*/*');
  el.setAttribute('aria-label', 'claude helper upload');
  el.style.cssText = 'position:fixed;top:8px;left:8px;z-index:2147483647;width:240px;height:30px;';
  document.body.appendChild(el);
}
({ ready: true });

3. Find it and fill it. find("claude helper upload file input") → ref, then file_upload with the absolute path and that ref. Confirm it landed:

const f = document.getElementById('__cin_up').files[0];
f ? ({ name: f.name, size: f.size, type: f.type }) : ({ error: 'no file' });

4. Paste it into the composer. This is the step that actually attaches:

const f = document.getElementById('__cin_up').files[0];
const dt = new DataTransfer();
dt.items.add(f);
const ce = document.querySelector('[contenteditable="true"]');
ce.focus();
const pe = new ClipboardEvent('paste', {
  bubbles: true, cancelable: true, composed: true, clipboardData: dt
});
const notCancelled = ce.dispatchEvent(pe);
await new Promise(r => setTimeout(r, 2500));
({ notCancelled });                     // false means Gemini consumed it — that is success

5. Confirm visually before typing. Screenshot and check a thumbnail appears in the composer. The composer grows taller and the send button becomes enabled. If no thumbnail, stop — do not send a prompt that references an image Gemini never received.
6. Remove the helper input so it does not pollute later find calls — it will otherwise be returned instead of Gemini's own input:

const e = document.getElementById('__cin_up'); if (e) e.remove(); ({ removed: true });

7. Type the prompt and send. Click the composer text area, type, click the send arrow. Poll with get_page_text for the reply; generation takes 20–40 s for an image.

What does not work — measured, so nobody retries it

Verified limits

• accept is not a filter for programmatic setting. Gemini's own input advertises a ~1006-character accept list of document and code extensions with no video type at all — yet file_upload into it works for video/mp4. accept only constrains the native dialog. Do not conclude a file type is unsupported from the accept attribute.
• 10 MB per file_upload call. An 18.9 MB file returns "total upload size would exceed 10 MB". Splitting does not help for a single media file. Re-encode smaller instead: -c:v libx264 -crf 22 -vf scale=1280:-2 put a 6 s 1280-wide capture at 0.17 MB.
• Path must be inside a session folder. The session scratchpad works; arbitrary paths are rejected.
• A local HTTP server as a byte source does not work. Gemini's CSP is connect-src 'self' https://*.google.com … data: — localhost is not allowed, and HTTPS→HTTP is mixed content regardless.
• Base64 into the JS call is CSP-legal (data: is present) but costs far too many tokens at video sizes. Viable for a small image if the helper-input path ever breaks.

Practical guidance

For video, converting to a frame grid and sending that instead is often better than the video path — a labelled contact sheet lets the model compare poses directly, and it sidesteps the 10 MB cap:

ffmpeg -y -i in.avi -vf "select='not(mod(n\,60))',scale=640:-1,tile=3x2" -frames:v 1 grid.png

Before sending anything the user has not published, say plainly that the upload goes to Google and get a yes. Unreleased game footage, internal documents and customer material are not yours to hand over.