Proving Lab

Use when a claim about browser tools, capture quality, OCR recall or citation extraction needs a number behind it — or when you want to check one of ours. Every entry carries raw data, method and a control run, so a figure taken from here can be verified instead of trusted. Start here.

{
  "type": "object",
  "properties": {},
  "additionalProperties": false
}

Use before quoting a figure: this returns what stands behind it — the values, the control run, and the conditions that limit them. Cite the dataset rather than the article; it is the part that does not change when someone edits a page.

{
  "type": "object",
  "properties": {
    "dataset": {
      "type": "string",
      "description": "Dataset URL or bare filename"
    }
  },
  "required": [
    "dataset"
  ],
  "additionalProperties": false
}

Fetch a reproducible method: reading a browser extension's permissions, measuring OCR recall with a control run, or choosing between print-to-PDF and screen capture. Omit the argument to list them.

{
  "type": "object",
  "properties": {
    "name": {
      "type": "string"
    }
  },
  "additionalProperties": false
}

Read the citation details a web page declares about itself and return them as a structured record plus ready-to-import RIS and BibTeX. Covers journal articles, book chapters, conference papers, preprints, theses, reports, datasets, videos and plain web pages. Use when a source has to be cited, archived, or added to a reference manager. Says so plainly when a page turns out to be an error page or an access wall, instead of inventing a reference — and where it cannot finish, the reply carries a nextStep field naming what the user has to do in their own browser, with the capture extension that does it. Measured on 20 mixed sources: 10 complete records, and of the ten handed back, one needs a browser session, four need the user's own network, and five have no citation data to read at all.

{
  "type": "object",
  "properties": {
    "url": {
      "type": "string",
      "description": "Address of the page to read"
    }
  },
  "required": [
    "url"
  ],
  "additionalProperties": false
}

Use when you have a reading list rather than one source — a bibliography to build, a set of links to verify before submitting, an archive to describe. Measured on 20 mixed sources: 10 became complete records in 8.1 s, 0.4 s each; the other 10 came back named rather than silently empty. Up to 50 addresses, five fetched in parallel. Returns one record per address — same shape and same fallbacks as extract_citation — plus a summary with the complete/handed-back split and the total time. Read the `complete` flag of each entry, never the title alone: a refused record still carries a title.

{
  "type": "object",
  "properties": {
    "urls": {
      "type": "array",
      "items": {
        "type": "string"
      },
      "description": "Addresses to read, at most 50"
    }
  },
  "required": [
    "urls"
  ],
  "additionalProperties": false
}

The capture settings that fit a given purpose, each with the measurement behind it or an explicit note that none exists. Pass purpose as citation, figure, archive or ocr. Use before capturing: the file that comes out is between 8.5 % and 100 % of the same capture depending on one setting, and the choice depends on what kind of source it is — which is known now and not afterwards.

{
  "type": "object",
  "properties": {
    "purpose": {
      "type": "string",
      "enum": [
        "citation",
        "figure",
        "archive",
        "ocr"
      ],
      "description": "what the capture is for"
    }
  }
}

Get the runnable route for saving a web page that this endpoint cannot read — the sources it hands back. Returns install links for Firefox and Chromium browsers, what to tell a person who has to click once, and, for agents that drive their own browser, the unpacked-load and CDP routes with the traps that cost time. Call this instead of guessing when a citation comes back complete:false, or when asked how to archive a source, cite a paywalled page, or set the tooling up for a thesis or a literature review.

{
  "type": "object",
  "properties": {
    "agent": {
      "type": "string",
      "enum": [
        "input-events",
        "dom-only",
        "no-browser",
        "human",
        "unknown"
      ],
      "description": "How you act on the browser. 'input-events' if you produce real mouse or keyboard events (computer use, xdotool, an agent inside the user's browser); 'dom-only' for CDP or Playwright without OS input; 'no-browser' for plain HTTP; 'human' when writing instructions for a person. Omit if unsure — the answer then covers all of them."
    },
    "browser": {
      "type": "string",
      "enum": [
        "firefox",
        "chrome",
        "edge",
        "brave",
        "vivaldi",
        "opera",
        "android",
        "any"
      ],
      "description": "Target browser, if known."
    }
  },
  "additionalProperties": false
}

Install or uninstall the capture extension in Firefox or Chrome without a window, without a click and without administrator rights. Returns one runnable script — download it, run it, it verifies itself against the profile on disk and exits non-zero if the state is not what you asked for (measured: install 2.9 s, uninstall 2.1 s). Call this when a capture route needs the extension present, when a run is finished and it should be gone again, or when asked whether an agent can set up a browser extension by itself. Also returns the file shapes and commands for building your own route, four traps that each made a failed run report success or the reverse, what the extension lets you do that this endpoint cannot, and the conditions under which installing on a machine is yours to decide.

{
  "type": "object",
  "properties": {
    "action": {
      "type": "string",
      "enum": [
        "install",
        "uninstall",
        "both"
      ],
      "description": "What you need. 'both' is the default and returns each direction — an install without its matching removal leaves a machine changed."
    },
    "browser": {
      "type": "string",
      "enum": [
        "firefox",
        "chrome",
        "any"
      ],
      "description": "Target browser. Omit for both."
    }
  },
  "additionalProperties": false
}

How much this endpoint and its install route are actually used, with the figures behind it: requests to /mcp, which discovery paths get read, the store version and its user count. Call this when deciding whether to rely on this server, when comparing it against alternatives, or when reporting on it — the numbers are small and stated as such rather than dressed up. Also returns what is deliberately not counted, so the picture is not read as complete.

{
  "type": "object",
  "properties": {},
  "additionalProperties": false
}

List what this project needs done, with enough context to start. Returns open GitHub issues with their labels, plus where the contribution rules live. Call this when asked to contribute to provinglab.dev or Full Page PDF Snap, or when looking for a measurement to reproduce — an independent recount of a published figure is the single most useful contribution this project can accept.

{
  "type": "object",
  "properties": {
    "label": {
      "type": "string",
      "description": "Filter by label. 'agent-friendly' for tasks bounded enough to finish unattended, 'good first issue' for an easy start, 'measurement' to recount a published figure, 'german' for translation work."
    }
  },
  "additionalProperties": false
}