Gemini 3.1 Flash Lite can return structured results and call tools through a compatible agent, but the model itself cannot see your signed-in browser. Select the stable Gemini API ID gemini-3.1-flash-lite, then separately confirm that your client can route requests to it and run local tools. Google’s API deprecation page lists May 7, 2027 as its shutdown date; Google Cloud model availability follows a separate lifecycle. Use Tabbit Browser with a local Tabbit CLI agent, in a named task tab, and review a saved draft before any submit action. Gemini API model catalog · Gemini 3.1 Flash Lite model page · Google Cloud listing
Verify the model and client separately
Google’s Gemini API catalog lists the stable gemini-3.1-flash-lite ID. The separate preview model uses a different ID and has a different lifecycle. The current API deprecations page lists May 7, 2027 as the shutdown date for the stable model, with Gemini 3.5 Flash Lite as its replacement. Recheck that page before publication.
The official model documentation lists function calling and structured outputs. Those capabilities still need a compatible agent client to connect model output to tools. The model does not gain your website login or browser tabs by itself. The Google Cloud Agent Platform model page is a separate product surface, so check its availability and lifecycle independently from Gemini API.
| Task | Route | Check first |
|---|---|---|
| Repeated lightweight extraction | Gemini API stable model | Exact ID and client availability |
| Read a signed-in webpage | Local agent + Tabbit CLI | Launcher and site session |
| Fill a form for review | Agent saves a named draft | Fields, destination, and source |
| Cloud-hosted task | Google Cloud Agent Platform | Its model ID and lifecycle separately |
Connect the model to a local browser agent
Load the Tabbit skill using your client’s own skill discovery syntax. Pi uses /skill:tabbit; some clients use /skills or automatically discover installed skills. Then check the Tabbit CLI launcher and the model provider route. Finally, confirm that the target site is signed in inside Tabbit Browser. Skill discovery, launcher reachability, model availability, and website login are separate checks.
Tabbit CLI runs in the existing Tabbit Browser environment and can reuse that site login when the account has access. Use a separate named task tab so your other work can stay open. A remote-only agent needs a separately supported local handoff; API availability alone does not give it local browser access.
| Layer | What it provides | What it does not prove |
|---|---|---|
| Gemini API model | Model responses and documented model capabilities | Access to your local browser |
| Agent client | Model routing and tool orchestration, if supported | That the target site is signed in |
| Tabbit CLI | Local browser control in Tabbit Browser | Permission to submit a form |
| Google Cloud listing | A separate managed model surface | Same lifecycle as Gemini API |
Check the exact API path
Two individual integration reports point to checks worth adding to your workflow. In a tau2-bench GitHub issue, an author reported that Gemini thought signatures were carried inside long tool-call IDs in their LiteLLM path; their correction says they did not test the failure after a transport change. A Google AI Developers Forum post describes completed Interactions API responses without a model output in that author's tests, while their tested generateContent calls returned outputs. These are specific reports, not general failure rates or evidence about Tabbit. Check the API route your agent actually uses and verify that a model output exists before transferring fields to a form.
Save a draft before submitting
Use a low-risk test form first. Ask the agent to find one named record on the signed-in page, extract only the needed fields, fill a draft, and stop. Give the task a clear name such as Gemini intake draft. The review checkpoint should show the source URL and each field beside its visible evidence. A person checks the account, destination, and values before submitting.
Prompt
On the currently signed-in page, find the record named [record name]. Return JSON with name, status, date, source_url, and visible evidence for each field. Fill the test form as a draft only. Do not click Submit. Use null for missing or ambiguous values and list any field that needs human review.This is an illustrative workflow, not a claim that a live form was filled. Do not place credentials, email addresses, or unrelated personal data in the prompt.
Related guides
Read the Tabbit CLI complete guide, learn how to connect an agent to an existing browser, follow the first browser-agent task checklist, or compare browser MCP and CLI. The Codex Tabbit CLI setup shows a client-specific setup. Also see reuse a login session, validate a local web app, and browser research with sources.
FAQ
What is the exact Gemini 3.1 Flash Lite API ID?
The stable Gemini API model ID is `gemini-3.1-flash-lite`. The preview model has a different ID.
Can Gemini 3.1 Flash Lite control my browser by itself?
No. A compatible agent must connect the model to local browser tools and the signed-in Tabbit environment.
Is the Gemini API model the same as the Cloud listing?
No. Google Cloud Agent Platform is a separate service. Verify its model status and lifecycle independently.
How do I load the Tabbit skill?
Use your client’s documented discovery method. Pi uses `/skill:tabbit`; some clients use `/skills` or automatic discovery.
How do I prevent an accidental form submission?
Ask the agent to save a named draft and stop before Submit. A person should check the destination, fields, and source before sending.