Perceptron Mk1 is a vision and video model, not a browser operator. Use it to interpret screenshots or clips, then hand a short, checked plan to a local agent that can run Tabbit CLI. The direct Perceptron API uses perceptron-mk1; OpenRouter routes use the namespaced ID perceptron/perceptron-mk1, and the OpenRouter page currently shows inconsistent Mk1/Mk1.5 capability text. Do not assume tool support from the model name; confirm the selected provider route, and keep Mk1 in its verified visual role until then. Keep Mk1 in the perception role and let a named local agent perform browser steps in a separate Tabbit task tab. Perceptron announcement · OpenRouter model page · OpenRouter Endpoints API · Perceptron API announcement
Use Mk1 for visual evidence
Perceptron describes Mk1 as a model for video and embodied reasoning. Its announcement covers image and video understanding, visual grounding, OCR, and structured annotations. That makes it a candidate for reading a visible status or finding a moment in a clip. It does not make the model a browser agent.
The OpenRouter Endpoints API for perceptron/perceptron-mk1 lists structured_outputs and omits tools and tool_choice. The model page has inconsistent Mk1/Mk1.5 text, so use the endpoint record for this route and keep direct Perceptron API support unverified. Structured spatial outputs such as points, boxes, polygons, or clip timestamps are still model output. They are not clicks. For signed-in web actions, use a separate client that can run a local browser tool.
| Task | Route | Check first |
|---|---|---|
| Read a screenshot or clip | Perceptron Mk1 | Confirm the model route accepts the media input |
| Navigate or fill a page | Local agent + Tabbit CLI | Confirm local tool access and website login |
| Interpret evidence, then act | Mk1 for reading; agent for actions | Verify each extracted field against the page |
| Use a remote-only agent | Supported local handoff, if available | Confirm the handoff feature exists |
Set up a local handoff
Load the Tabbit skill with your agent client's own skill discovery method. Pi uses /skill:tabbit; some clients use /skills or discover installed skills automatically. Skill discovery only loads instructions. It does not confirm that the Tabbit CLI launcher works, that the selected host can call perceptron/perceptron-mk1, or that the target website is signed in. Check these separately.
Tabbit CLI operates in the existing Tabbit Browser environment and can reuse the site login there, subject to the account's access. Use a separate, named task tab so you can keep working in your other tabs. A remote runner needs a separately supported local handoff; model availability alone does not grant access to a local browser.
| Capability | Perceptron Mk1 | Browser workflow |
|---|---|---|
| Image and video understanding | Documented focus | Useful for reading visual evidence |
| Structured spatial annotations | Explicit output modes | Data for review, not an action |
| Function calling on OpenRouter route | Not available | Do not ask this route to click |
| Website session access | Not supplied by the model | Use Tabbit CLI in the local browser |
Keep the example small and reviewable
Start with an illustrative test page, not a consequential form. Ask Mk1 to read a release-note screenshot. Then have the local agent open that page, compare the visible title and version, and save a draft called Mk1 release-note draft. The draft should show the source URL, extracted fields, and any uncertainty. A person checks the result before it is shared.
This is a sample workflow, not a claim that Mk1 or Tabbit performed a live test. If a value is unreadable or the page differs from the screenshot, stop and ask for review instead of guessing.
Prompt
Review this screenshot as evidence for a release-note draft. Return JSON with visible_title, visible_version, exact_visible_status, evidence_text, confidence (0 to 1), and needs_human_check (boolean). Do not infer hidden page content. Return null for any unreadable field.Related guides
Read the Tabbit CLI complete guide, learn how to connect an agent to an existing browser, follow the first browser-agent task checklist, or compare browser MCP and CLI. The Codex Tabbit CLI setup covers a client-specific example. See also reuse a login session, validate a local web app, and research with sources.
FAQ
Can Perceptron Mk1 click buttons?
The OpenRouter route `perceptron/perceptron-mk1` does not support function calling. Use a separate local agent for browser actions.
Which model ID should I configure?
OpenRouter lists `perceptron/perceptron-mk1`. Confirm your chosen agent client can route requests to that exact model before starting.
Does structured output mean the model can operate a browser?
No. Structured annotations describe the model response. They do not provide browser controls or a signed-in session.
Can Tabbit CLI reuse my signed-in website session?
Tabbit CLI runs in the existing Tabbit Browser environment, where it can reuse a site's login if that account has access.
How should I review a browser task?
Save a named draft with its source URL and extracted fields. Check each field against the page and pause before submit or publish for human review.