To use Muse Spark 1.1 with a signed-in browser, route the exact ID muse-spark-1.1 through Meta Model API in a compatible local agent, then connect that client to Tabbit CLI. Meta still lists 1.1 on its API but says it is unavailable in Muse Code; its docs recommend 1.3 for new work. Check access before building around the older ID, then compare both routes on the same harmless test page and leave a named draft for review. Meta's version table · Tabbit CLI quick start
Confirm the exact route before opening the browser
| Need | Route | Verify first |
|---|---|---|
| Reproduce an existing 1.1 workflow | Meta Model API, muse-spark-1.1 | Confirm account access and record the returned model ID |
| Start new Muse Spark work | Compare the API's recommended muse-spark-1.3 | Check the client supports the exact route you will evaluate |
| Use the 1.1 model in Muse Code | Not available per Meta's version table | Choose Model API with a compatible client or use a supported local handoff |
| Inspect a signed-in page | Local agent plus Tabbit CLI | Verify skill discovery, launcher reachability, profile, and site permissions separately |
Meta lists three hosted API IDs: muse-spark-1.3, muse-spark-1.2, and muse-spark-1.1. Its current docs call 1.3 the latest and recommend it for new work; 1.1 remains Standard mode only and does not appear in Muse Code availability. This is hosted inference, not a model installed on your laptop. A model directory entry also does not prove your client account or local tools can use it. Check Meta's model catalog, coding agent setup, and API overview for the route you actually have.
Keep skill discovery and browser access separate
Install Tabbit Browser and sign into the test site yourself. In the local agent, use its documented skill discovery method. Pi uses /skill:tabbit; some clients use /skills, while others load installed skill files automatically. Separately verify that the local tabbit-cli launcher runs. Then open the authorized test page in the Tabbit profile where you signed in. Passing one check says nothing about the others.
Use a separate task tab so your work stays available. Tabbit CLI operates in the Tabbit Browser environment and can reuse that site's signed-in session; it does not sign into Meta or grant the page permission. Do not copy cookies, tokens, or account data into the model prompt. If the agent is remote-only, hand off to a local agent only through a client-supported mechanism. Otherwise have an authorized operator run the prepared steps locally.
Compare the versions on a harmless test page
Choose a page you own or can access, with a stable heading and one visible status field. Ask the local agent to make a read-only pass with the exact 1.1 route, then repeat with 1.3 if both routes are available. Keep the prompt, page, account, and extraction fields identical. Compare missing values and source URLs, not a vague impression that one answer “looks smarter.” This is a user-run comparison recipe, not a benchmark result from this article.
Use the installed Tabbit skill from this local agent. Open a separate task tab to my authorized test page. First confirm the configured model ID is exactly `muse-spark-1.1`; if the provider cannot confirm it, stop. Read only the visible page title, status, and owner. Return the page URL and exact visible text for each field. Do not click controls or follow other links. Prepare a draft named "Muse Spark 1.1 test-page read — DEMO-217" with the model ID, timestamp, fields, source URL, and unknowns. Do not save page changes or publish the draft. Show it for review.
Expected output: verified model ID, page URL, title, status, owner, unknowns, and a reviewable draft link.DEMO-217 is an illustrative draft name, not a real workspace item. For a comparison, repeat the same steps with 1.3 and place both outputs side by side. A successful text call alone does not prove browser access or correct page reading.
Decide from the observable result
| Result | Next step |
|---|---|
Provider rejects muse-spark-1.1 | Stop and verify account, endpoint, and exact ID; do not silently accept an alias |
| 1.1 works through API but is absent from Muse Code | Use a compatible API client or evaluate a currently supported route |
| Client cannot call local tools | Use a documented local handoff or have an operator run the checklist |
| Page redirects to login or shows another account | Sign in yourself or stop and confirm the intended profile |
| One output has missing or unsupported fields | Keep that field unknown; review the cited page before any action |
For browser setup, see the Tabbit CLI guide, existing-browser connection guide, first task checklist, and MCP vs CLI guide. The Codex setup, agent browser setup, browser automation overview, browser tools guide, and browser security notes cover related choices. Read Meta's Muse Spark version guide before choosing between 1.1 and 1.3.
FAQ
Is Muse Spark 1.1 still available?
Meta's current Model API docs list `muse-spark-1.1` as a Standard-mode API ID. The same page says 1.1 is not available in Muse Code and recommends 1.3 for new work.
Can Muse Spark 1.1 directly control Tabbit Browser?
No direct Tabbit connection is documented. A compatible local agent must call the hosted model route and run Tabbit CLI as a separate local tool.
Can I use Muse Spark 1.1 in Muse Code?
No. Meta's version availability table marks Muse Spark 1.1 as unavailable in Muse Code. Use Model API only if your account and provider route permit the exact ID.
Can the local agent reuse my website login?
The local agent can use the signed-in Tabbit Browser environment through Tabbit CLI when the page is authorized. The model does not receive login credentials or site permission.
Should I migrate to Muse Spark 1.3?
Meta recommends 1.3 for new work and describes it as tuned for agentic workflows. Compare the exact task in your own test page and pin the returned model ID before deciding.
What if my client cannot run local tools?
Use an explicit local handoff only if the client supports it. Otherwise ask it to prepare the page checklist and have an authorized local operator perform the browser steps.