If you can serve Laguna XS.2 locally and your agent client can run local tools, connect that agent to Tabbit CLI to read an authorized dashboard in Tabbit Browser. XS.2 is still downloadable under Apache 2.0, but it is no longer Poolside's newest XS model, and the hosted Poolside Chat/API shutdown date has passed. Treat the model endpoint, tool calling, browser connection, and website login as separate checks.
Key takeaways
Laguna XS.2's official weights remain available on Hugging Face under Apache 2.0. Poolside now lists XS 2.1 as the newer XS release.
Poolside's current Chat page says service ended on October 31, 2026. An old OpenRouter model page is not proof that a request will route today.
Tabbit Browser holds the signed-in website session. Tabbit CLI lets a local agent work in that browser. Laguna XS.2 only supplies model output through the inference endpoint you configure.
Start with a read-only dashboard check. Return changed rows, source URLs, unknowns, and a review note. Do not let the test edit the source.
Check how the model will be served
The release post describes Laguna XS.2 as a 33B-total, 3B-active mixture-of-experts model, released under Apache 2.0 on April 28, 2026. It documents local use with Ollama and gives ollama launch pool --model laguna-xs.2. Poolside later noted a 256K served context update on May 26, 2026. Those are model and launch facts, not a promise that your hardware, runtime, or agent client will work without setup. (Poolside release · XS.2 model card)
At the time of this research, the live serving route needs extra care. Poolside Chat says it shut down on October 31 and points users to host the weights themselves. The OpenRouter pages still show descriptions and stale free-tier wording, but a November 10 check of OpenRouter's public model catalog returned no XS.2 model entry. Do not paste an old model ID into an agent and assume it works. Recheck the provider's live model list and run a harmless request before exposing any private page. (Poolside Chat notice · OpenRouter model page · OpenRouter model catalog)
| Route | What is verified | What you must verify next |
|---|---|---|
| Local weights | The official model card remains available; Apache 2.0 is listed. | Download the exact variant and check the license, runtime, quant, memory and context settings. |
| Local inference | Poolside documented Ollama launch instructions at release. | Confirm your installed runtime still supports this model build and returns tool calls in the format your agent expects. |
| Hosted API | Poolside's prior Chat/API shutdown date has passed. OpenRouter's page remains visible, but the public catalog check had no XS.2 route. | Do not rely on a hosted route until the current provider catalog and a test request both confirm it. |
| Browser access | The separate Tabbit CLI fixture test verified session reuse in its own browser environment. | Confirm the local agent can launch the CLI and that Tabbit Browser has the intended account open. |
Keep the four connections separate
There are four links in this workflow. The local inference runtime serves XS.2. An agent client selects that model and decides when to call tools. The Tabbit skill tells the client how to use Tabbit CLI. Tabbit Browser supplies the page and its current site session. A working link in one layer does not establish the next.
Install/open Tabbit Browser on the same machine where the agent will call Tabbit CLI. Install the Tabbit skill in the client using that client's documented skill discovery. Clients that support named skills may expose /tabbit; Pi documents /skill:tabbit. Some clients use /skills or discover skills from a folder without a slash command. Check the client's actual interface and make sure the skill launcher can reach the local CLI. Then separately confirm the site itself is signed in inside the selected Tabbit profile. (Tabbit CLI quick start · Connect an agent to an existing browser · Codex setup)
| Check | Pass condition | If it fails |
|---|---|---|
| Model route | A harmless completion returns from the exact XS.2 endpoint. | Verify the endpoint still serves XS.2; do not silently substitute XS 2.1. |
| Tool call | The client receives and parses the expected tool call. | Check model template, serving runtime, parser and client support with a harmless local task. |
| CLI reachability | The client discovers the Tabbit skill and can run Tabbit CLI. | Fix local skill discovery or launcher permissions before troubleshooting the model. |
| Site session | The selected Tabbit Browser tab shows the intended signed-in dashboard. | Sign in manually in that browser profile or choose the right profile; never copy cookies into the prompt. |
Read a dashboard without changing it
Use a dashboard where you are authorized to read. The sample name Release review — GEO-104 and its board rows are illustrative fixtures, not a test performed with XS.2. Ask the agent to open the board in a named task tab, compare visible status and modified date with your expected baseline, and return links for each row. Do not ask it to press Save, Resolve, Publish or another mutating control.
Use Laguna XS.2 only through the configured local inference endpoint. In the current Tabbit Browser profile, open the dashboard I name: [authorized dashboard URL].
Read only. Compare the visible rows with this baseline: [row IDs and expected status/date]. Return a table with row ID, visible status, visible modified date, source URL, baseline difference, and confidence. Quote no private text beyond the fields listed. Mark missing or conflicting values as unknown. Do not click buttons that save, resolve, publish, delete, invite, or change data. Stop after producing a source-linked report named "Release review — GEO-104" in the separate task tab. This name is an example, not a real dashboard record.If a client understands skills by name, invoke /tabbit before giving the work request; in Pi, use /skill:tabbit. If the client only offers /skills, open its skill list and select Tabbit there. Do not assume a slash command exists just because another client uses one. Read the browser agent first-task checklist before a task involving sensitive or irreversible controls.
Put Tabbit Browser between the agent and the dashboard
Tabbit is useful here because the human can sign in to the intended profile first, while the local task stays in a separate tab. The model does not receive the browser's cookies as prompt text. The agent asks Tabbit CLI to read pages inside the browser environment, then returns a report for review. This boundary depends on the agent and CLI running locally in that environment; a cloud agent does not gain access to the browser just by choosing Laguna XS.2.
The parent task recorded a local fixture test: the first named CLI task reached a protected page and received 401 before sign-in. After a short-lived test cookie was set, a second named task reached the same fixture page, saw the heading “Release notes review” and ID GEO-104, and received HTTP 200. This is evidence of session reuse in a test fixture. It is not an XS.2 run, external agent integration, or test against a real dashboard. The example above is a proposed workflow only. (Tabbit CLI guide · Browser profiles and task tabs · What an agentic browser does)
Compare the report with what the page shows
The report is ready for review only when each row links to the same dashboard record you can open yourself. Reopen the source link, compare the visible status and date, and keep ambiguous changes out of the summary. A successful tool call proves that a call returned; it does not prove the agent selected the right board, read every row, or interpreted a change correctly.
| Report state | Evidence to open | Reviewer action |
|---|---|---|
| Matched | Same row ID and visible status/date on the linked page. | Keep the finding and its source URL. |
| Changed | Current page differs from the supplied baseline. | Confirm the baseline date and record the exact fields that changed. |
| Unknown | Row is absent, hidden, or ambiguous. | Leave it unresolved and ask the owner; do not fill the gap from memory. |
| Conflicting | Page and model summary disagree. | Trust the visible source after checking the correct account and page. Correct or remove the report row. |
| Ready for review | All reported values can be reopened; source pages remain unchanged. | A person decides whether a later update is authorized. |
Community reports are useful as cautions, not capability tests. One LocalLLaMA commenter said an early XS.2 run “was pretty good” but noted a 131K context limit in that setup. Poolside later documented a 256K served-context update, so the practical lesson is to verify the runtime, quant and actual limit you selected. A separate creator, Fahd Mirza, documented installing and testing XS.2 locally with vLLM; the video page does not establish browser use or a repeatable result. (Reddit discussion · Fahd Mirza's local test)
DannyTup ran a separate XS.2 benchmark on DGX Spark and reported lower results than Qwen in that setup. His comparison is a reason to repeat your own fixed-input test before using the model for dashboard reads. Original NVIDIA forum post
For a disconnected task, stop and identify which link failed: model endpoint, tool parser, local CLI, or signed-in page. Resume with a read-only check after fixing that one link. The MCP versus CLI guide helps when choosing the local connection; the source-linked research workflow explains how to review evidence.
If you already have the exact XS.2 weights and a compatible local runtime, try the read-only recipe on a low-risk dashboard. Keep the final report separate from the source until you have reopened and checked every link.
FAQ
Can Laguna XS.2 open my signed-in browser by itself?
No. Model weights do not include your browser cookies or tabs. A local agent client must call a running Laguna XS.2 endpoint and invoke Tabbit CLI in the Tabbit Browser environment where you signed in.
Is Laguna XS.2 still available as a hosted API?
Poolside Chat says the service shut down on October 31, 2026. The OpenRouter page still renders an old listing, but the model was absent from its public catalog check on November 10, 2026. Verify a live route before configuring a client.
Is Laguna XS.2 still the latest Poolside XS model?
No. Poolside's official model pages list Laguna XS 2.1 as the newer XS release. XS.2 remains downloadable, and its model card lists Apache 2.0.
Can I use Tabbit CLI with a model hosted on another machine?
The agent must be able to run the Tabbit skill and local CLI where Tabbit Browser is available. A remote model endpoint may be callable by that local agent, but a remote runner does not automatically inherit the browser session.
Will the example change dashboard data?
No. The prompt asks for a read-only comparison and a source-linked report. Keep the agent from saving edits and verify every reported row in the browser.
What should I check if tool calls fail?
Check the model endpoint and runtime's tool-call parser first, then check that the client discovered the Tabbit skill and can launch Tabbit CLI. Finally confirm the correct Tabbit profile is signed in to the intended site.