Agent Stack Shorts: Muse in Chrome, Poppy, and Free Models Worth Testing
Muse in Chrome, Poppy, Claude Code billing fixes, Antigravity workflows, and three free models worth testing.
🎧 Listen to EpisodeAgent Stack Shorts: Muse in Chrome, Poppy, and Free Models Worth Testing
November 14, 2026 · English · 5:16
A compact Agent Stack Shorts edition in the AgentStack Daily feed. Muse's Chrome extension and Poppy offer two different ways for agents to work with websites. This edition also covers a Claude Code billing fix, Antigravity workflows, and three free model routes.
Muse in Chrome: a different browser connection
Meta's extension works in your own Chrome, using existing signed-in sessions and a dedicated tab group. That can remove a login obstacle caused by a separate cloud browser. It does not establish that a website's restrictions on automation have disappeared.
Test the actual workflow in stages: reading the page, navigating results, and reaching the requested action. A successful page load alone does not demonstrate successful shopping or checkout. The audio discusses mixed early Amazon reports as reports, not as our own successful test.
Meta's browser and Chrome-extension guide
Poppy 1.0: discover supported actions rather than guess
Sierra's November 13 announcement describes a protocol for agents to discover how a company supports interaction through websites, APIs and business agents. Sierra says Muse checks for Poppy first and Shopify publishes support for merchants. This creates a route to an explicitly supported capability instead of relying only on clicking through an interface. It is not a claim of universal checkout or confirmed Amazon participation.
Sierra's release announcement · Protocol specification
Claude Code 2.1.325: check the billing path
The November 13 release fixes background sessions launched with a Claude login using an API credential inherited from the starting shell. It also repairs Code Mode worktree and path handling. If you use both subscription access and API credentials, update the client and verify which route your background job uses. These fixes are not an increase in Claude Pro usage allowance.
Antigravity: workflows versus model entitlement
App 2.32 adds Dynamic Workflows with a bundled Python runtime for parallel and multistep agent work. Try a bounded implementation-and-review task, with distinct jobs for each agent. The app rollout is gradual. CLI options for Hydrogen and Carbon with API-key authentication should not be confused with Google AI Pro subscription inclusion.
Free routes to try
- Ling 3.2 Flash: text and image input, roughly one million tokens of context, tool calling and structured-output support. A useful first test is a public screenshot converted into a structured record, followed by a small coding task with clear tests. OpenRouter listing
- Recraft V4.1 Flash V2: text-to-image generation, intended for rapid new-image iteration rather than reference-image editing. The free listing is scheduled to end November 28, 2026. Check any lettering carefully. OpenRouter listing
- Inkling 2 Small: multimodal, tool-using reasoning model; the free route is for agentic harnesses. Its 17 billion active parameters do not equal its total 594 billion parameters, so the name and active count do not establish local hardware fit. The audio makes no claim that it runs on a single DGX Spark. OpenRouter listing
These are separate hosted model routes, not additional allowance in existing coding subscriptions. Availability and free-route limits can change. The practical tests in this edition are deliberately bounded so results can be compared against tools you already use.
Claude Code has a fix worth checking before your next coding session. Anthropic released version two point one point three two five on November thirteenth. It fixes background sessions started with a Claude login unexpectedly using an API key inherited from the shell. If you keep both subscription access and API credentials around, that distinction matters: a background job should not quietly move from your subscription to API billing. The release also fixes Code Mode using the wrong worktree and certain paths landing inside the project. These are client repairs, not an increase in Claude Pro's allowance.
Google's Antigravity update adds another way to coordinate agents. App version two point thirty-two adds Dynamic Workflows, using the bundled Python runtime for parallel agents and multistep work. A practical experiment would be to have one agent implement a small change and another check it. The app update rolls out gradually. Separately, the command-line release adds Hydrogen and Carbon options for API-key authentication. That is not the same thing as Carbon being included in Google AI Pro. Keep API access separate from subscription access when choosing the model.
Two new free OpenRouter listings are especially useful this week. Ling three point two Flash appeared on November thirteenth. It accepts text and images, supports tool calling and structured output, and lists a context window of just over one million tokens. Try giving it a public screenshot and asking for a structured record, or compare it with your usual model on a small coding task with clear tests. This is a free OpenRouter route, not a new benefit inside your existing subscriptions.
The second is Recraft V four point one Flash V two. Its free endpoint also appeared November thirteenth, and is currently listed to end November twenty-eighth. It generates images from text, but does not accept reference images. That makes it useful for fresh podcast-cover concepts, not for matching or editing an existing cover. The listing advertises improved text rendering. Check the lettering, title and episode number yourself before using the result. Both free routes are rate-limited.
Inkling two Small is another model worth revisiting. Its free OpenRouter route dates to November ninth, and Thinking Machines confirmed downloadable weights on November twelfth. The hosted route supports image inputs and tool calling, but does not enforce JSON response format. It is restricted to agentic harnesses. Despite the name Small, the model has five hundred ninety-four billion total parameters, with seventeen billion active. That is not enough information to claim it will fit on a single DGX Spark. Local fit still needs to be tested.
The useful detail about Muse's Chrome extension is how it connects to your browser. Meta's help page says it uses your existing Chrome browser and logged-in sessions, and works inside its own tab group. That means a task can start where you already have the account open, instead of making you sign in again inside a separate cloud browser. For a workflow that gets stuck at the login screen, that could remove a real obstacle.
But a working login and a working website automation are different things. Reading a product page is one test. Clicking through search results, opening the cart, and completing checkout are separate tests. Reports about Amazon are mixed, including a tester who initially reported success and then corrected it after Amazon blocked the first click. So the extension changes where Muse works, but it does not establish that Amazon's agent restrictions are gone. The useful comparison is the same task in the existing Chrome session: can Muse read it, can it navigate it, and can it complete the requested action? That is the difference between a browser connection that looks promising and one that actually fixes the workflow.
There is also a new route beyond clicking through a website. Sierra announced Poppy one point zero on November thirteenth. Poppy gives an agent a way to discover which capabilities a site supports. Sierra says Muse checks for Poppy first, and Shopify automatically publishes a Poppy file for merchants. That could let Muse find a supported way to perform a task instead of guessing its way through the interface. But capability discovery is not a promise of universal checkout, and the announcement does not establish Amazon participation. For shopping, the stronger signal is a merchant explicitly exposing the action the agent needs, not simply the fact that the agent can open the page.