Granite 4.2 8B can work with signed-in pages when a local agent connects Tabbit CLI to the browser session. The hosted API or local model does not receive browser cookies. Check the model route first, then use a separate task tab and review a named, unshared draft. The model card lists Apache 2.0, 8B parameters, a 128K context window, and vLLM/SGLang serving paths. This is an open-weight local-serving route; the cited material does not establish a hosted IBM API for this exact checkpoint. Tool execution still depends on the serving parser and agent. See the publisher model documentation and client or product setup.
Key takeaways
Verify the exact model route, provider and region before starting.
Skill discovery, CLI reachability and website login are three separate checks.
Keep the first task read-only and limited to authorized pages.
Ask for URL-backed facts in a named draft; review before sharing.
Choose the route
| Setup | What it supplies | Decision |
|---|---|---|
| Model weights | Apache 2.0 8B checkpoint | Review license and hardware |
| Serving | vLLM or SGLang OpenAI-compatible endpoint | Confirm tool-call parser |
| Browser access | Local Tabbit CLI through an agent | Keep execution on signed-in machine |
For Qwen Code, skill discovery uses ~/.agents/skills/ and the /skills or /skill:tabbit command. For other clients, use their documented skill invocation; do not assume a /tabbit command. In any client, seeing the skill does not prove the CLI can launch, and a working CLI does not prove the target site is signed in. Install Tabbit Browser, sign in yourself, then use a separate task tab while keeping your other tabs available.
Prepare a source-linked draft
Use ibm-granite/granite-4.2-8b in this local session. Use Tabbit CLI to read only the three team pages I authorize. Create a draft named "Granite 4.2 8B source brief, draft 01". Add the exact URL and a short evidence excerpt for every factual claim. Separate fact from inference, mark missing information unknown, and do not edit source pages. Stop before sharing or publishing and show me the draft for review.Expected output: a short draft with source URL and evidence excerpt per factual point, explicit unknowns, and no submitted changes. This prompt is illustrative; it is not a result from a model run.
Review before sharing
| Check | Pass | Stop |
|---|---|---|
| Server | Endpoint responds with intended model | Wrong checkpoint or parser |
| Browser | Authorized profile in task tab | Login screen or wrong account |
| Tool scope | Read-only, three named pages | Unbounded shell or navigation |
| Draft | Named, cited, unshared | External submission |
What community reports do and do not show
One Reddit user reported running Granite 4.2 8B Q8 with Pi and llama.cpp, and described repeated tool-call reasoning loops during a simple Flask CLI edit. The same author separately raised path-scope concerns in an MCP shell setup. Treat these as one person’s setup reports, not independent evidence or a model-wide failure. Read the firsthand discussion and related setup report.
Use Tabbit as the local page layer
Install Tabbit Browser. Sign in to the target site in that browser profile and ask your local agent to use Tabbit CLI. A fixture check for this batch confirmed that a second named CLI task could reuse the signed-in session and read the authorized local test page. It did not call an external model or test this model’s integration. The task tab stays separate so you can continue working in your other tabs. Keep the draft unshared until you verify each source.
Continue with the Tabbit CLI guide, connecting an agent to an existing browser, session reuse, the first-task checklist, MCP vs CLI, profiles and task tabs, Codex setup, and browser-agent security.
FAQ
What model ID should I use?
Use ibm-granite/granite-4.2-8b and verify that your local server loaded this exact checkpoint. Hosted availability depends on the provider.
Can Granite open my signed-in pages by itself?
No. A local agent must connect Tabbit CLI to the browser session. The model itself does not receive browser cookies.
How do I confirm browser tool calls work?
Check the loaded checkpoint and serving parser, then verify that the agent can call Tabbit CLI from the same machine and profile.
What should the first task do?
Read a few authorized pages, attach URLs and evidence to each claim, and save a named draft for review.
Was Granite 4.2 8B tested with Tabbit?
No model-to-Tabbit run was completed for this draft. The local fixture verified only the browser session-reuse path.