To use Mistral Small 4 with a signed-in website, choose a client that can select the documented mistral-small-2603 route and run local tools, then connect Tabbit CLI to Tabbit Browser. Read a bounded dashboard view and return a named draft such as “Weekly changes — review draft”; keep the save or publish action with a person. Mistral’s official docs identify the model as Apache 2.0, but that does not guarantee that your chosen client exposes it.
Key takeaways
Model access, tool calling, local launcher access, and the browser session are four separate checks.
The documented API ID is
mistral-small-2603; confirm it in the provider and client you will actually use.A local agent can read a signed-in Tabbit Browser session through Tabbit CLI. Keep its first task read-only.
Use a named draft and a visible review checkpoint. The workflow does not imply a native Mistral–Tabbit integration.
Check the route before opening a browser
| Check | Evidence to look for | Stop when |
|---|---|---|
| Model route | Provider model list includes mistral-small-2603 | The ID is missing or account access is unavailable |
| Local tools | Agent client can launch the installed Tabbit CLI | Only a remote runner is available, with no local handoff |
| Tool-call format | Client and provider accept the tool schema used by the agent | A harmless tool call fails or is silently ignored |
| Browser profile | Tabbit Browser has the intended site session | The page asks you to sign in or shows a different account |
Mistral’s model page lists mistral-small-2603, weights, function calling, and an Apache 2.0 license. The release announcement describes Small 4 as a 119B-parameter mixture-of-experts model with image input. “Open weights” does not mean a small download or a lightweight local run. Its serving requirements depend on quantization and runtime. The API path and a locally hosted checkpoint are different deployments.
A Reddit user describing a Mistral Small 4 + Hermes + Open WebUI setup said, “Multi-step tool use actually holds together, the agent loop is transparent”. That is one user’s report about a particular harness. It does not establish browser compatibility in your client. Read the thread.
Use a read-only dashboard task
First sign into the target website in the Tabbit Browser profile you intend to use. Open a separate Tabbit CLI task from the local agent. Ask it to inspect one page, collect a few visible fields, and stop. Skill discovery and executable access are different: a client may show /tabbit or /skills while still lacking permission to start the CLI. Pi uses /skill:tabbit. Follow the syntax exposed by your own client.
Try this prompt with a test dashboard or a low-risk view. It asks for concrete output and forbids writes.
Read the “Operations overview” page in the signed-in Tabbit Browser session through Tabbit CLI. Do not click controls that change data, send messages, or publish. Return a draft titled “Weekly changes — review draft” with: (1) each metric label and its visible current value, (2) changes explicitly shown on the page, (3) items that need a human check, and (4) the source URL beside every claim. If a value is not visible, write “not shown”. Stop before saving anything.Review the URL and each value in the browser. If the page shows a login prompt, wrong workspace, or incomplete content, stop instead of asking the model to infer missing data. Keep the draft in the agent response or a clearly named scratch document until you accept it. The same-page replay documented for the Tabbit CLI checked that a second task could read a page in the same signed-in browser environment; it did not test Mistral Small 4 or prove an external agent integration.
Decide what to verify after the first read
| Situation | Next check | What it does not prove |
|---|---|---|
| Text fields appear complete | Compare each value with the visible page | That the model is right about hidden or stale data |
| A chart or image matters | Confirm the selected model route supports image input and test one harmless image | That the browser tool exposes screenshots to the model |
| Tool call is rejected | Check client schema, provider route, and local permissions separately | That the model has no general tool-use ability |
| You need a saved report | Move the reviewed text into the destination yourself | That the agent should have write or publish access |
Put the workflow into Tabbit Browser
Tabbit Browser is a practical option when the task depends on a site where you are already signed in. Install it, sign in to the intended account, and use Tabbit CLI from a local agent to open a separate task space. This lets you keep working in your other tabs while the agent reads the requested page. Review the named draft in the browser before copying it into a shared report.
This is a user-configured workflow, not native model compatibility. The model provider controls model access; the agent client controls tool execution; Tabbit CLI connects the local task to the browser. A cloud-hosted agent needs a separately supported local handoff. For setup, see the Tabbit CLI guide, existing browser connection guide, first-task checklist, CLI versus MCP guide, Codex setup guide, agentic browser explainer, browser agent guide, and Tabbit Browser overview.
Sources
FAQ
Does Mistral Small 4 open a browser by itself?
No. A local agent client must connect browser tools, and the model route must support the tool-call format that client uses.
What is the current API model ID?
Mistral documents mistral-small-2603. Confirm the ID in your provider and client before configuring the workflow.
Can I run the weights locally?
Mistral publishes weights under Apache 2.0. The model is large, so check your inference hardware and serving stack before downloading.
Can Tabbit reuse a website login?
A local Tabbit CLI task can use the Tabbit Browser environment where you are already signed in. Keep credentials out of prompts.
Does every client have a /tabbit command?
No. Skill discovery differs by client. Some use /skills; Pi uses /skill:tabbit. Check that client’s own help.
Can a remote runner reach my desktop browser?
Only if that setup supports a separate local handoff. A cloud model endpoint alone does not gain access to a desktop browser.