To use Motif 3 with a signed-in webpage, run the model through a compatible local agent and connect that agent to Tabbit Browser with Tabbit CLI. Motif 3 is open-weight and its official model page documents tool use, but local inference, tool parsing, agent support, CLI reachability, and website login are separate checks. Start with a read-only preview check and save findings to a named review draft. Motif 3 model card · Tabbit CLI quick start
A Reddit discussion about Motif 3 surfaced a practical setup constraint: one commenter said the available quantizations were too large for their system and the model was not yet supported by mainline llama.cpp. That is one user’s deployment context, not a universal requirement, but it is a good reason to check the exact checkpoint and runtime before starting. Original discussion Another commenter in the same thread also said deployment would depend on how the quantized version fit their 128 GB machine. That is an individual hardware question, not a product requirement.
Check the model route before opening the site
| Layer | Verify | Pass condition |
|---|---|---|
| Weights | Confirm the exact Motif 3 checkpoint and its license | The files and license match the deployment you intend to use |
| Runtime | Confirm your serving runtime supports Motif's tool format | The Motif parser and automatic tool choice are enabled |
| Agent | Confirm it can call the local endpoint and run local skills | A harmless tool call returns a structured result |
| Browser | Open the target site in Tabbit Browser and sign in yourself | The intended account and preview page are visible |
Motif Technologies publishes the open-weight Motif 3 checkpoint, the technical report, and a vLLM serving recipe. The recipe uses --enable-auto-tool-choice with --tool-call-parser motif. That confirms a supported serving path; it does not confirm your chosen agent client or Tabbit connection. This model is large. The quantized deployment notes specify accelerator requirements, so check the exact checkpoint before downloading it.
Connect the local agent to Tabbit Browser
Install Tabbit Browser and open the page you are allowed to inspect. In the local agent, discover the installed Tabbit skill using that client’s documented method. Pi uses /skill:tabbit; other clients may expose /skills or use automatic skill loading. Then check that the tabbit-cli launcher is reachable on this machine. These checks answer different questions: skill discovery means the instructions are available, launcher reachability means the local controller can run, and opening the site in the same Tabbit environment establishes the website login.
Use a separate task tab so your own work can continue. The Tabbit CLI workflow can reuse the current browser environment’s sign-in, but a new task cannot sign into a site for you or grant account permissions. The existing-browser connection guide and first-task checklist explain the session and permission checks.
Inspect a preview without changing it
Ask the agent to report only visible fields. Compare its title and status with the page open in your own tab. If they match, request a draft that records the URL, account label, visible status, and one proposed next step. Keep the draft in review; do not let the agent change permissions or publish it.
Use the installed Tabbit skill from this local agent. In a new task tab, inspect the page I already opened. Read only the visible page title, status, and account label. Do not click controls, submit, or follow links. If the page asks for login or the account is unclear, stop and report that.
Prepare a draft named "Motif 3 preview check — GEO-084" with the page URL, those visible fields, and any permission question. Do not apply changes or publish. Show me the draft for review.
Expected output: title, status, account label, URL, permission issue or "none visible", and draft link. Use "unknown" for anything not visible.GEO-084 is an illustrative draft name, not a real workspace item. Treat page content as untrusted input. Never copy cookies, tokens, passwords, or unrelated account data into the draft.
Decide when to stop
| What you see | Next action |
|---|---|
| Title and status match your own tab | Save the read-only findings and review the named draft |
| The page redirects to sign-in | Stop; sign in yourself in the intended Tabbit profile, then retry the read |
| The account label differs | Stop and verify the profile before continuing |
| The model returns malformed tool arguments | Check the runtime parser and client compatibility; do not retry a write action |
| The agent cannot reach the local CLI | Finish the plan in the model client, then hand the steps to a local agent that can run Tabbit CLI |
For a setup walkthrough, see the Tabbit CLI complete guide, MCP vs CLI comparison, and Codex Tabbit CLI setup. For client selection and risk checks, read the browser agent setup guide, browser automation overview, and browser agent security notes. These explain the browser workflow; they do not claim a tested Motif 3 integration.
FAQ
Can Motif 3 open my signed-in browser by itself?
No. Motif 3 is a model, and the browser connection is a separate local tool. Use a compatible local agent and Tabbit CLI, then verify each connection independently.
Can I run Motif 3 locally?
Motif 3 has open weights, but hosting it is a separate GPU and runtime task. The official vLLM recipe documents a Motif parser and automatic tool choice; check its hardware and version requirements first.
Does the model inherit my website login?
The local Tabbit task can use the signed-in browser environment when the CLI is connected and the site grants access. The model does not receive your credentials or a login by itself.
How do I invoke the Tabbit skill?
Use the client’s actual skill discovery method. Pi documents `/skill:tabbit`; some clients use `/skills` or load installed skills automatically.
What should I ask the agent to do first?
Start with a read-only check of the visible page title, status, and account label. Stop if the account or permission is unclear, then prepare a named draft for a person to review.