TabbitBlog

Use Gemma 4 12B IT with a signed-in browser

Connect Gemma 4 12B IT to a local browser agent, verify function calling in your runtime, and review authenticated page work safely.

In this article
  1. Key takeaways
  2. Connect the browser through a local agent
  3. Worked example: verify a support ticket without changing it
  4. Decision: use the model for a bounded read-only browser task
  5. Continue with the browser workflow

To check a signed-in support ticket with Gemma 4 12B IT, install Tabbit Browser and use Tabbit CLI from a local agent. Confirm that its runtime supports the exact checkpoint and structured tool calls. Sign in yourself, then let the agent collect the ticket ID, status and source link in a separate task tab for review.

Key takeaways

  • Gemma 4 12B IT is open-weight and Google documents function calling, but the chosen runtime must implement it.

  • A model response is not a browser action. The local agent must expose tools, the runner must parse calls, and Tabbit CLI must reach the intended browser.

  • Check model and runtime versions, parser support, and a harmless tool-call fixture before opening a real account.

  • Tabbit Browser can keep the existing site login in its target environment while a task uses a separate tab.

  • Treat page text as untrusted input; pause before submitting, publishing, buying, or changing account settings.

CheckPass conditionIf it fails
Model routeProvider/runtime resolves the exact Gemma 4 12B IT revisionVerify model ID, license acceptance, and provider availability
Tool callingA harmless declared function produces a structured call and the host executes it onceCheck chat template, tool schema, parser, and runtime adapter
Browser handoffLocal agent invokes the installed Tabbit CLI and identifies the intended instanceCheck skill discovery and launcher reachability separately
LoginTarget site is already signed in inside the intended Tabbit environmentSign in manually; never send credentials to the model
ReviewReport is saved and a person checks it before any write actionKeep task read-only and stop at the review checkpoint

Connect the browser through a local agent

Install the model in a runtime that explicitly lists the Gemma 4 12B IT checkpoint and tool/function calling. Check the actual provider’s current model ID, access, context limits, and tool support; an open Hugging Face weight file does not mean every hosted API offers it. Google’s model card and developer guide describe the model and function calling.

Next, confirm the agent client can discover the installed Tabbit skill and run its launcher locally. Those are separate checks: a skill file being present does not prove the client loads it, and a working CLI does not prove a page is signed in. In clients that support named skills, use their documented /skills selector; in Pi, invoke /skill:tabbit. Do not assume a /tabbit command exists in other clients.

Open the target site in the intended Tabbit environment and sign in there yourself. A CLI task can reuse that environment’s existing login, while its owned task tab stays separate from your other work. A remote model runner needs a separately supported local handoff; network access to the public site is not access to your local browser.

Worked example: verify a support ticket without changing it

Ask the local agent to open the support queue, read one visible ticket, and save a draft report named gemma-ticket-check-104.md. The report should contain the ticket’s visible ID, title, status, source URL, and any mismatch. Do not include email addresses, message text, cookies, or other unnecessary personal data. Keep the task read-only.

A useful instruction is:

Read the visible support ticket in the signed-in Tabbit task tab. Do not submit, edit, reply, or change account settings. Save `gemma-ticket-check-104.md` with the visible ticket ID, title, status, URL, and a short verification note. If the page is ambiguous or access is denied, stop and report that. Do not copy personal data.
SituationRecommended ownerStop point
Public docsModel/runtime or local agentConfirm source and version
Signed-in ticket or dashboardLocal agent plus Tabbit CLISave a minimal report
Draft wordingModel in the clientReview the proposed text
Reply, submit, publish, or account changeHumanConfirm exact target and content before action

Decision: use the model for a bounded read-only browser task

Gemma 4 12B IT is a reasonable local candidate when your chosen runtime supports the exact checkpoint and structured function calling. Tabbit CLI provides the browser handoff; it does not make the model itself a browser controller. Validate the chain with a harmless fixture first, then use a named task and a visible human review step. If the runtime cannot emit calls the host understands, use the model to draft instructions and have a supported local agent perform the browser work.

Continue with the browser workflow

Tabbit CLI setup · connect a local agent · first browser-agent task · CLI versus browser MCP · Codex setup · reuse a signed-in session · browser session basics · local web app checks · permission boundaries

FAQ

Does Gemma 4 12B IT support tool calling?

Google documents native structured function calling for Gemma 4. The runtime and agent must still support the checkpoint, template, parser, and tool execution path.

Can it control a browser by itself?

No. A local agent must expose browser tools and a runner must execute model calls. Tabbit CLI is a separate browser handoff.

Can I run it through any hosted API?

No. Check the provider’s current model ID, access, region, context limit, and tool support. Open weights do not imply universal hosted availability.

Can it reuse my website login?

The model does not handle credentials. A local Tabbit CLI task can use the existing login in the intended Tabbit environment after you sign in there.

What should I test before using a real account?

Test a harmless declared function, verify the host executes it once, and then run a read-only browser check with a named report and human review.

Take the next step

Let Tabbit work alongside you.

Research across tabs, automate repetitive browser work, and keep every piece of context within reach.