Use Schematron V2 Turbo with a Signed-In Browser

Schematron V2 Turbo cannot operate a signed-in browser by itself. It is a schema-guided HTML extractor, and Inference.net explicitly says its Schematron extraction models do not support tool calling. To prepare a team-document draft, use a separate local agent or a person to open authorized pages through Tabbit Browser and Tabbit CLI, pass selected HTML to Schematron for structured extraction, then have a human review the named draft against the source.

The distinction is practical: inference-net/schematron-v2-turbo returns fields that match a schema; it does not navigate, click, or write. The official Schematron guide says the schema itself supplies the extraction instructions, without natural-language system or user prompts. The coding-agent documentation lists the Schematron extraction family among models without tool-calling support.

Key takeaways

• Use Schematron V2 Turbo to extract typed fields from HTML, not as the browser-controlling agent.

• The exact Inference.net model ID is inference-net/schematron-v2-turbo.

• A local operator must open the signed-in page. Tabbit CLI can provide that browser step when the chosen client can run local tools.

• Put extraction guidance in the JSON Schema field descriptions. The model does not take regular system or user prompts.

• Keep internal source pages read-only. A human should compare every extracted value with the page before saving a shared draft.

Split the browser and extraction work

This separation also helps with failure diagnosis. Tabbit's /tabbit skill (in clients that support that invocation) teaches an agent how to use its local launcher. Some clients provide a skill picker instead; Codex users can check the /skills list. Skill discovery, CLI reachability, and the website's login are different checks. Read how to connect an agent to an existing browser before using a work profile.

Read only the pages you are allowed to use

Install Tabbit Browser and open the team page yourself in the intended profile. Confirm your organization allows the page content to be sent to the Inference.net API before sending HTML. Then have a local agent use the installed Tabbit skill and CLI in a separately named task tab. If your client cannot run local tools, hand the exact page URL and read-only checklist to a local agent or a person. A remote model does not inherit your browser's session.

For example, an operator might collect release-note headings and visible status from two approved team pages. Keep this first pass read-only. Ask the operator for the source URL and page title alongside the text so you can compare the extraction with the open tabs. The first browser-agent task checklist covers that kind of bounded read, and web search versus browser control explains why a search snippet cannot confirm private page access.

The model's job begins after the HTML is available. Define fields clearly and make the schema do the work of the prompt.

import { readFile } from "node:fs/promises";
import OpenAI from "openai";
import { z } from "zod";
import { zodResponseFormat } from "openai/helpers/zod";

const client = new OpenAI({
  baseURL: "https://api.inference.net/v1",
  apiKey: process.env.INFERENCE_API_KEY,
});

const ReleaseNote = z.object({
  title: z.string().describe("Exact visible title of the release note."),
  updatedDate: z.string().describe("Visible update date, or 'not shown'."),
  status: z.string().describe("Status text shown on the page, or 'not shown'."),
  owner: z.string().describe("Named team or owner shown on the page, or 'not shown'."),
});

const sourceUrl = new URL(process.argv[2]);
if (!["http:", "https:"].includes(sourceUrl.protocol)) {
  throw new Error("Expected an HTTP or HTTPS source URL");
}
const html = await readFile("approved-record.html", "utf8");

const response = await client.chat.completions.parse({
  model: "inference-net/schematron-v2-turbo",
  messages: [{ role: "user", content: html }],
  response_format: zodResponseFormat(ReleaseNote, "release_note"),
  temperature: 0,
});

const extracted = response.choices[0].message.parsed;
if (!extracted) throw new Error("No parsed record returned");
console.log(JSON.stringify({ ...extracted, sourceUrl: sourceUrl.href }));

Save the approved HTML as approved-record.html. Pass the URL captured by Tabbit as the script’s first argument. The client attaches that URL after extraction, so the model only supplies the four page fields.

Convert extracted fields into a review draft

Schematron's JSON is input to a separate writing step. The browser operator or a general-purpose local agent can organize the extracted values, but it must not imply that the extractor itself can open the browser or update the team workspace.

Use the extracted fields and their source URLs to prepare a draft named
"Release notes review - November 10".
For each page, include the title, visible update date, status, owner,
source URL, and any field marked "not shown".
Do not open write controls, save, or submit a change to the team document.
I will compare every field with the original pages in my regular Tabbit tabs
before the document owner decides what to save.

The named draft gives the document owner a concrete review checkpoint. Keep the original source pages open in your regular Tabbit tabs while a separate task tab reads them. A local synthetic session test in this environment showed a first request returning 401 and a later Tabbit task reading the fixture after a temporary cookie was set. The fixture used no team account and the test did not call Schematron or change a document.

Diagnose by capability

The local client step is separate from model selection. Tabbit's CLI supports browser tasks through compatible local clients, but this article does not claim that Schematron is a supported Tabbit agent model. See browser MCP versus CLI for the connection options, and the Tabbit CLI guide for local task reads. What an agentic browser does explains the browser's role. Install Tabbit from the download page if you need a local browser environment.

We found no qualifying first-hand community report about V2 Turbo operating a signed-in browser. One older Reddit thread contains a user report about rate limits and server errors in the earlier Schematron API generation. It is not evidence about V2 Turbo, so it is not used to describe current behavior. The workflow here is an illustrative handoff, not a tested model integration.

Use the model for the extraction job it documents. Let a local operator handle the authorized browser read, then compare a named draft with each source page before anything is saved.

Secholar.L

Growth & Content at Tabbit

Secholar.L writes hands-on guides about AI browsers and agent workflows, comparing tools with real research and citing verifiable sources.