Perceptron Mk1.5 can interpret images, video, and audio and can call functions, but that does not give it access to a desktop browser session. To review a signed-in page, run a local agent on the same machine as Tabbit Browser, use Tabbit CLI to open the page, and pass only a redacted screenshot or extracted text to the model. The agent then returns a named review draft with a human checkpoint.
This is a local-client, model-route, and browser-connector workflow. It is not a claim of native Tabbit integration. Perceptron’s Mk1.5 release notes say the model accepts text, image, video, and audio, and supports JSON-Schema function tools. OpenRouter currently lists perceptron/perceptron-mk1.5 with tool calling and one serving provider. Check your own account’s route and region before building around it.
Key takeaways
The model interprets page content and prepares a draft; a local agent calls Tabbit CLI.
Check the model route, skill discovery, launcher connection, and site login separately.
Keep the first run read-only, include source URLs, and stop at a human review checkpoint.
OpenCode's
@tabbitsyntax is client-specific, not a universal command.
Check model and browser access
| Layer | What to check | Pass condition |
|---|---|---|
| Model | Official or API provider for Perceptron Mk1.5 | The model appears in your current account picker and tool settings are documented |
| Client | OpenCode or another host with local tools | It discovers the skill and can call the model route |
| Skill | ~/.agents/skills/tabbit/SKILL.md, invoked as @tabbit in OpenCode | The client confirms the skill loaded |
| Launcher | ~/.local/bin/tabbit-cli | It diagnoses and reaches the intended Tabbit environment |
| Website | The page you are authorized to read | You signed in within the Tabbit Browser environment the CLI reaches |
Check each layer separately. A model that answers a prompt does not prove it has browser tools. An open Tabbit tab does not prove the site session works. Configure Perceptron Mk1.5 in a client that exposes the OpenRouter route and can run local shell tools. OpenCode can discover the installed Tabbit skill from ~/.agents/skills/tabbit/SKILL.md; invoke it as @tabbit in an OpenCode prompt, then verify that the skill is actually loaded. This is client skill discovery. Separately check that ~/.local/bin/tabbit-cli launches and reaches the intended Tabbit instance. Finally, confirm the website itself is already signed in there. One passing check does not prove the next one.
Choose text extraction or visual handoff
This recipe is for checking a permission preview, account badge, or status label that is clearer in the rendered page than in its text. Use a non-sensitive test page first. The model never receives the browser login itself.
| Page evidence | What to pass to the model | When it fits | Control |
|---|---|---|---|
| Readable text and tables | DOM or OCR text extracted by the local agent | Labels, dates, and field values | Remove unrelated personal data; keep the URL |
| Layout or badge state | A cropped, redacted screenshot | Meaning depends on visual position or icon | Confirm the client supports image attachments |
| Ambiguous or conflicting state | A small text excerpt and one crop | One evidence type is inconclusive | Flag for review instead of guessing |
If the page is mostly text and labels, use DOM text extraction and skip image upload. If the meaning depends on layout, icon state, or a visual badge, pass a cropped, redacted screenshot through the local agent only after confirming the client supports image attachments. Do not assume the model’s visual input is the same as browser control.
Create a named draft and stop for review
Sign in to the site yourself in Tabbit Browser. Then ask your local agent to read the named pages, use the model to organize evidence, and save a separate draft. The names and rows below are an illustrative test fixture, not proof of access to your workspace.
Use the local Tabbit skill through Tabbit CLI to open the authorized preview page in Tabbit Browser. Read the page title and capture only the visible status panel. Redact names, email addresses, tokens, and unrelated account details before sending a screenshot or OCR text to Perceptron Mk1.5. Create a draft named "Access preview review, draft 01" with columns: visible label, exact page URL, observed status, confidence, and question for a human. Separate what is visible from your interpretation. Do not click approval, change permissions, or submit. Stop and show the draft and the redacted evidence. If image transfer to the model is unavailable, use OCR text from the local agent and state that no image was analyzed.Expected output: A small table with one row per visible status, each tied to the page URL, followed by a "Needs human review" list. No write actions.
Open the saved draft in a separate Tabbit Browser task tab, reopen every source link, and leave the document marked "Needs human review". You can keep using your other browser tabs while this review task runs.
| Draft status | Meaning | Next step |
|---|---|---|
| Supported | The page directly shows the field or state | Reopen the source and confirm it |
| Unclear | Related text exists but does not settle the claim | Keep it under "Needs human review" |
| Missing | The target field or record is absent | Do not fill it in automatically |
| Ready for review | Every claim has a URL and no write action ran | Review the full draft before deciding |
Install Tabbit Browser and connect the local agent
Start by installing Tabbit Browser, then sign in to sites you are authorized to use. The local Tabbit CLI entry on macOS/Linux is ~/.local/bin/tabbit-cli; an installed skill can live at ~/.agents/skills/tabbit/SKILL.md. OpenCode can invoke a discovered skill with @tabbit. Other clients may expose a skills list or use another syntax, so do not assume a universal /tabbit command.
Skill discovery, CLI access to the browser, and the site's existing login are separate checks. In a local test, a second named task reused a short-lived test session in the same Tabbit environment. That verifies session reuse in that tested environment; it does not prove native integration with Perceptron Mk1.5. Task tabs are separate, so you can keep using other browser tabs while the agent works. For the connection boundaries, read Tabbit CLI guide, connect an agent to an existing browser, the first-task checklist, and the browser MCP vs CLI guide. For client-specific setup, see Codex CLI setup. For the browser itself, install Tabbit Browser; review browser-agent security and the browser-for-agent overview before granting tools.
Use visual handoff for a read-only review
In the University of Basel test archive, Perceptron Mk1.5 scored differently on extraction and transcription. Test the exact screenshot or text format you plan to send, then compare the extracted status with the original page. Original test archive
Based on the official material, Perceptron Mk1.5's documented input and tool abilities can help organize evidence on the model side. A local Tabbit agent still opens the signed-in page. If your client cannot load the skill or send images, use text extraction and label it accurately; do not describe an unavailable capability as tested.
Try the read-only recipe on a non-sensitive test page first. Save it as "Perceptron Mk1.5 browser review, draft 01", verify every URL and status, then decide whether the same recipe fits a real workspace.
FAQ
Can Perceptron Mk1.5 open my signed-in Tabbit tabs by itself?
No. It accepts multimodal input and supports function calls, but it does not inherit a local browser session. A local agent must use Tabbit CLI to capture the page or extract its text.
Can the model read a screenshot?
Its provider listing documents image input. Your chosen agent client must also support sending an image to that provider. If it cannot, use local OCR or DOM text and say which evidence was used.
Which API model ID should I try?
Perceptron lists `perceptron-mk1.5` for its own API. OpenRouter lists `perceptron/perceptron-mk1.5`. Verify the live model picker, region, and account before relying on either route.
How does the agent find the Tabbit skill?
In OpenCode, put the skill in a supported discovery directory such as `~/.agents/skills/tabbit/SKILL.md` and invoke `@tabbit`. Confirm discovery in that client; the launcher and website login are separate checks.
Can I let the model approve access?
Keep this flow read-only. Use the draft to identify visible states, then have an authorized person open the source page and decide whether to approve anything.
What should the review draft contain?
Record the page URL, exact visible label, extracted or visual evidence, and an uncertainty field. Name it clearly, for example `Access preview review, draft 01`, and stop before any write.