Hy3 Preview can request a browser tool call, but it does not automatically inherit a Tabbit Browser login. The model, the agent client, and the browser runtime are separate parts. To compare two pages behind a login, install Tabbit Browser, set up its local CLI in a supported agent client, and use only an account you are authorized to access. Save the comparison as a named draft and leave edits or submissions for a person.
The April preview route is currently listed by OpenRouter as tencent/hy3-preview. Tencent’s July production Hy3 is a separate release. Do not apply the later release’s Apache 2.0 terms to the preview: the preview model card supplies its own community license and excludes the EU, UK, and South Korea. Check the exact checkpoint, provider, account, and local legal requirements before sending data. Tencent preview license, OpenRouter provider list
What the tool call means
A model may return a structured request to open a page, search, or extract fields. A compatible client decides whether that request is allowed, runs the browser operation, and sends the result back. Tabbit CLI is the local bridge between a supported agent client and Tabbit Browser; it does not make a remote model endpoint a browser session. OpenCode exposes the discovered skill as @tabbit; Pi uses /skill:tabbit. For other clients, inspect the client’s skill list rather than assuming a universal command.
| Part | Responsibility | Check |
|---|---|---|
| Hy3 Preview | Plans and returns a tool request | Confirm exact route and model terms |
| Agent client | Applies policy and invokes a tool | Confirm it can use Tabbit CLI |
| Tabbit Browser + CLI | Opens pages in your local signed-in session | Use an authorized account and a separate task tab |
| Person | Verifies and approves the result | Review source pages before any write or submit |
A bounded comparison workflow
Install Tabbit Browser and complete local CLI setup using the current Tabbit CLI guide. Confirm the client can see the Tabbit skill and the browser reports a ready connection.
Start a separate, named task tab, such as
Record comparison / HYP-203 / draft 01. Use a signed-in account only if you are authorized to view both records. Do not pass passwords, cookies, or session tokens to the model.Give the agent two specific page URLs and a read-only task. Ask it to compare named fields, quote short evidence, include each source URL, and report missing or conflicting values. Do not ask it to edit, send, or submit.
Reopen both pages yourself. Check the values and citations, then save the output as a draft ending in
Needs human review. Keep the pages and draft separate from any production record.
Example prompt
Using the authorized signed-in Tabbit Browser session, open only these two pages: [record A] and [record B]. Read the fields ID, owner, status, and due date. Return a comparison table with the value from each page, a short visible-text excerpt, and the source URL. Mark missing or conflicting values as “Needs human review.” Do not edit, click submit, send messages, or follow links outside these pages. Save a separate draft named “Record comparison / HYP-203 / draft 01” and stop.This is an illustrative prompt, not a claim that Hy3 Preview was tested with Tabbit. Public reports about Tencent’s later Hy3 using Tabbit describe a different production model and client setup; they are not evidence about this April preview. Likewise, a model’s advertised tool-call support does not prove the runtime can execute a browser action. Test the exact provider and client with non-sensitive pages first.
Keep the handoff reviewable
A useful draft includes the two page titles, access date, URLs, requested fields, extracted values, and unresolved differences. Avoid including unrelated account data. If the agent cannot reach Tabbit, stop the cloud run and hand off locally: copy only the research question and public references into a local client connected to Tabbit CLI. Do not expose a login to a cloud agent or create a remote tunnel just to bridge the browser.
| Field | Check on both pages | Put in the draft |
|---|---|---|
| Record ID | Compare the full visible ID before reading other fields | Report mismatch and stop if they differ |
| Owner and status | Copy each page’s current labels | Keep values and URLs in separate columns |
| Due date | Check timezone or displayed format | Mark any missing value for review |
See the existing-browser connection guide, the first-task checklist, and the browser MCP versus CLI comparison for setup details. For a broader overview, read agentic reasoning and deep research.
For setup context, see the Tabbit Browser installation walkthrough, Codex and Tabbit CLI setup, browser agent security risks, and choosing a browser for agents.
FAQ
Can Hy3 Preview control Tabbit Browser directly?
Not by itself. Hy3 Preview can produce tool-call requests, but a compatible client and runtime must execute them. A local agent can use Tabbit CLI if configured; the model endpoint is not a browser connection.
What is the exact Hy3 Preview model route?
The current OpenRouter listing uses `tencent/hy3-preview`. Check provider availability, account access, and regional terms before use; routes can change.
Does the preview execute its own browser tools?
No. Tool calling describes structured requests. The client or serving runtime executes the browser operation and returns results to the model.
Is Hy3 Preview the same as the later production Hy3?
No. The public preview released in April is distinct from Tencent’s later production Hy3 release in July, which Tencent announced under Apache 2.0. Verify the license attached to the exact checkpoint you deploy.
What should I check before using the model?
Review the model card, license, territory restrictions, serving requirements, and provider terms. The preview license excludes the EU, UK, and South Korea.
How should I save browser research?
Create a separate named draft with page titles, dates, source URLs, and discrepancies. Reopen the pages and let a person approve any change or submission.