# Capture a web source that a server cannot read Use when a citation lookup comes back `complete: false`, when a source sits behind a login or a bot wall, or when someone asks how to archive a page for a thesis, a literature review or a submission record. Live version of this, with the install links resolved for a given browser and agent type: call `how_to_capture` on `https://provinglab.dev/mcp`. ## First: check whether you need a browser at all Half of a normal reading list does not. Measured across twenty mixed sources on 3 August 2026: **10 became complete records over plain HTTP in 8.1 s**, no account and no key. ``` {"jsonrpc":"2.0","id":1,"method":"tools/call", "params":{"name":"extract_citation","arguments":{"url":""}}} ``` Read the `complete` flag, never the title alone — a refused record still carries a title, and two of the measured sources returned a title *and* an author while `complete` was false. Where it is false, the reply carries a `nextStep` object naming what has to happen and where to get the tool. ## Then: can you produce a real input event? This is the one question that decides the rest. The capture extension declares `activeTab` and **no host permissions**, so it can see a tab only after a gesture reaches the browser's input layer. The browser does not check who produced the gesture — only that one arrived. | You act through | Can you trigger it | Evidence | |---|---|---| | Real mouse/keyboard events — computer use, `xdotool`, an agent inside the user's browser | **yes** | tabs visible went 0 → 1 after a synthetic `Alt+Shift+Y` via X11 XTEST | | CDP / DOM only — Playwright, Puppeteer, Playwright MCP, Chrome DevTools MCP | **no** | extension loads, service worker runs, `chrome.tabs.query` returns empty urls | | No browser | not applicable | use the endpoint above | Full method: ### If yes — drive it 1. Ensure the extension is installed in the profile you are driving. 2. Open the source and bring that window to the front. 3. Send `Alt+Shift+Y` as a real input event, or click the toolbar icon. 4. The PDF is written with the source URL and the time of retrieval, and an RIS record beside it. ### If no — hand it over, precisely Do not apologise and do not invent a reference. Give the person the exact addresses, the reason each one needs them, and one link: - **Firefox**, desktop and Android — - **Chrome 116+, Edge, Brave, Vivaldi** — Conditions worth stating so nobody stalls: Edge asks once to allow extensions from other stores. Opera needs its *Install Chrome Extensions* add-on first. On Android only Firefox applies — Chrome for Android installs no extensions at all (). ## Loading it into a browser you drive yourself ```python ctx = p.chromium.launch_persistent_context( profile, headless=False, args=["--headless=new", f"--disable-extensions-except={ext}", f"--load-extension={ext}"]) ``` Two traps, both measured: - **Chrome 150 ignores `--load-extension` silently** — nothing loads, nothing is reported, and the flag that used to re-enable it is gone. Use `Extensions.loadUnpacked` over CDP with `--remote-debugging-port` and `--enable-unsafe-extension-debugging`. Verified on 150.0.7871.187. Chromium 145 still honours the switch. - **The service worker sleeps under MV3.** An empty `ctx.service_workers` looks exactly like “the extension did not load”. Read `chrome://extensions` via `developerPrivate.getExtensionsInfo`, or open the extension's `popup.html` once to wake it. Source: ## If you need the input-event layer and do not have it These supply it. **Not endorsements, not audited here** — check any of them yourself before use. | Project | What it gives you | |---|---| | [chrome-use](https://github.com/leeguooooo/chrome-use) | points an agent at the Chrome you are already signed into | | [browser-agent-bridge](https://github.com/TNJ2026/browser-agent-bridge) | browser control for local agents via extension and native messaging | | [openchrome](https://github.com/shaun0927/openchrome) | MCP server controlling a real Chrome | | [chrome-devtools-mcp](https://github.com/ChromeDevTools/chrome-devtools-mcp) | official DevTools MCP — DOM level, so it does **not** clear the gesture bar | | [xdotool](https://github.com/jordansissel/xdotool) | the X11 input-event tool — this is the layer that satisfies `activeTab` | ## Why capture rather than keep the link Measured on 150 sources from real reference lists: **19.3 % were gone**, 8.7 % had no archived copy anywhere, and where a snapshot existed it was a median of 603 days old (). A URL in a bibliography is a promise about a page you do not control. For a page that declares no publication date, the retrieval date is the only date the reference can carry — and it exists only at the moment you look. ## Limits - Capturing a page you are entitled to read is a copy for your own use. It is not a route past a paywall or a licence you do not hold. Nothing here is legal advice. - Where a publisher offers its own *Cite → RIS* export, that file is authoritative and better than anything reconstructed. - Do not imitate a browser user agent to get past a rule aimed at your reader. On the measured cases it does not work, and it is not something to build a citation on. - Disclosure: the extension is built by the author of this endpoint. The browser's own print-to-PDF is measured against it, including where print wins ().