Use Laguna S 2.1 with a signed-in browser

To use Laguna S 2.1 with a signed-in browser, run it behind a tool-call-capable local endpoint, connect a local agent to Tabbit Browser with Tabbit CLI, and ask the agent to collect a few authorized sources into a separate draft. Stop before editing the source workspace. A Reddit user testing Laguna in an OpenCode loop reported that generation sometimes stopped before finishing a task; that is one setup, but it is a good reason to verify one short tool cycle before handing over a multi-page job. (Original report)

Poolside describes Laguna S 2.1 as a 118B model with 8B active parameters, built for agentic coding and longer tasks. Its model card documents tool-call parser options for vLLM and SGLang. Those facts make it possible to evaluate the model in a local agent workflow. They do not establish that it can draft reliable team updates, or that a particular agent client supports the model endpoint. (Poolside's release · Official model card)

Key takeaways

• Laguna S 2.1 is the model. A serving runtime exposes its endpoint, a local agent connects model and tools, and Tabbit CLI reaches the browser.

• Poolside's vLLM example includes the poolside_v1 tool-call parser and automatic tool choice. Check version and client support before relying on it.

• Tabbit CLI can use a signed-in session in the Tabbit Browser environment. It creates a task tab so you can keep working in your other tabs.

• Keep source reading read-only. Ask for a separate named draft with URLs, visible dates, and a human review stop.

• Treat the sample workflow as a test plan, not proof that Laguna has been evaluated on office documents.

Decide whether the connection is ready

Do not collapse these checks into one. Poolside's model card shows how to serve the checkpoint and configure tool calls. It does not select an agent client for you. The Tabbit CLI quick start covers the browser connector; the existing-browser guide explains the session boundary. For the sequence of checks, use the first-task checklist. If you are choosing between a browser MCP and a local CLI, see how the two connection types differ.

Collect sources before writing a shared update

Use a small page list, such as the latest internal release note, the project decision log, and a relevant status page. Record the exact title and URL from each page. Ask the model to mark a statement as observed only when the page visibly supports it; label conclusions as interpretations. The output should be a new note, not a rewrite of a shared source.

This keeps the model from turning an old note into a current status. It also gives the reviewer a compact way to find a mismatch. If the page is a project system with filters, read the browser profile and task tab guide before asking the agent to move between workspaces. For a public-source-only task, the source-linked browser research guide describes a lighter workflow. The agentic browser overview explains which part Tabbit Browser plays.

Use Tabbit Browser for a named, reviewable draft

Sign in to the pages yourself in Tabbit Browser. Start a local agent whose endpoint is confirmed to support Laguna S 2.1 and tool calls. In supported clients, the Tabbit skill is commonly invoked as /tabbit; first check the client's skill picker or run its setup check. The command is not a universal model feature. In the Tabbit environment, the local CLI opens a separate task tab that uses the browser session you already have there. The official CLI quick start explains the setup. Tabbit's Codex setup guide shows one client-specific configuration; do not assume every custom model route behaves the same way.

The following is an illustrative fixture prompt. Replace the sample names and URLs with pages you are allowed to read. It does not report a Laguna test result.

Use the Tabbit skill in a new task named LAGUNA-WEEKLY-UPDATE-01.
Open only the three project pages I provide in my signed-in Tabbit Browser.
Do not edit, submit, share, or change those pages.
Create a separate document named "Weekly engineering update, draft 01".
For each claim, include the source page title, exact URL, visible date,
and a short supporting excerpt. Mark each item observed, inferred, missing,
or conflicting. Draft no more than five update bullets from supported facts.
Stop and show the named draft with its source links for human review.
If a page is inaccessible, a tool call fails, or two sources disagree, stop
and report the last verified step. Do not switch models without asking.

Expected output is a short draft with a link beside every factual claim, plus a list of unresolved items. Open the draft in the separate task tab, follow each source link, and confirm the author has not changed the original pages. A local Tabbit CLI test in this environment showed that a second named task could reuse a short-lived test login and read a protected fixture page. That test did not use Laguna S 2.1, an external agent, or a team workspace.

Check the output before anyone updates the source

An early Laguna user reported that the model was thorough in research and reasoning on their own setup, while another OpenCode test report said its generation stopped before completing tasks. The contrast is useful as a reminder, not as a verdict: the reports used different machines, runtimes, and prompts. (Research-loop discussion · OpenCode test)

If Laguna stops mid-generation, retry only the unfinished read after checking the runtime's output and tool-call state. Do not replay a write action. If the model route is inaccessible, the browser session is missing, or the output contradicts the source, save a note about the failure and stop. The browser troubleshooting guide is a better next step than switching endpoints blindly.

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.