Kimi K3.1 Pro can help plan a browser task, but the model alone cannot reach a signed-in site. Choose an agent client that can call a Kimi K3.1 Pro endpoint and run local tools, then give that client Tabbit CLI. Tabbit Browser supplies the existing site session; you review each proposed change before it is submitted. For this walkthrough, the task is to find one invoice in two admin portals and prepare a reconciliation note, not to pay or edit the source records.
A practical starting route is OpenCode with the Kimi K3.1 Pro model listed by OpenRouter. The route exists, and Moonshot publishes the model weights, but that does not prove every client/provider combination supports the same tool-call format. Run a small read-only check before using it on a real account. Kimi Code and Kimi WebBridge are separate products and workflows. Neither is Tabbit CLI.
For background on the model itself, read the Kimi K3.1 Pro overview.
Key takeaways
Kimi is the planner and language model. The local agent client decides which tools can run.
Tabbit CLI lets a supported local agent operate pages in the target Tabbit Browser environment, where you have already signed in.
Keep lookup and drafting separate from actions that change billing, permissions, or records.
A model card or provider route does not validate your whole client setup. Confirm tool calls in a harmless test first.
Kimi K3.1 Pro browser use at a glance
| Need | What to choose | What to verify |
|---|---|---|
| Model access | Kimi K3.1 Pro through a listed API route, such as OpenRouter | The endpoint is enabled for your account and the client shows the intended model ID |
| Local orchestration | A desktop agent client that can use that endpoint and execute local tools | Tool calls work; the process runs on your machine |
| Browser access | Tabbit Browser plus the Tabbit CLI skill | The skill is discovered and the CLI can reach the intended Tabbit instance |
| Website access | Sign in to the relevant admin pages in that Tabbit environment | The page is already authenticated before the agent starts |
| Review | A named draft in a separate task tab | Compare the draft against both source records before any save or submit |
The Kimi model card identifies an open-weight model and shows self-hosting routes such as vLLM. OpenRouter lists a hosted model route. Those are different ways to provide inference. Neither setting connects the model to your browser. See the Kimi model card and OpenRouter route.
Prepare the client and browser separately
Install and open Tabbit Browser. Sign in to the two admin sites yourself, using the Tabbit environment you intend the agent to use.
Configure your local agent client to use a Kimi K3.1 Pro endpoint. OpenRouter lists the
moonshotai/kimi-k3.1-proroute. Check your account's current model list and tool-call support before you configure the client.Install or expose the Tabbit skill to that client. In clients that support the named skill command, invoke
/tabbit; in Pi, the documented form is/skill:tabbit. Some clients discover skill files through/skillsor load them automatically. Follow that client's own skill instructions instead of assuming the command works everywhere.Ask the agent to run a read-only page-title and heading check. Confirm the CLI is using the intended Tabbit instance and that the target portal shows the account you signed in to.
This checks three separate links: model access, local tool execution, and browser-session reachability. A successful chat response only checks the first.
Use a reconciliation draft with a human checkpoint
A safe task is to compare one invoice reference across two signed-in admin pages and prepare a note. Give the agent a narrow scope, allow read-only browsing, and ask it to stop before making a record change.
Use Tabbit CLI to open the billing portal and vendor portal that are already signed in.
Find invoice INV-4827 in both. Compare vendor, invoice date, currency, and total.
Do not edit either portal, approve payment, or submit a form.
Create a draft named "INV-4827 reconciliation (review)" with a table of the values,
source page for each value, and any mismatch. Stop and show me the draft for review.
If a page asks for sign-in or blocks access, stop and report which page failed.The exact syntax depends on the client. Its skill may be a slash command, auto-discovered file, or a tool instruction; the task text is an example, not a tested Kimi-agent integration. Tabbit CLI runs locally, and task tabs are separate from the tabs you use to continue working. The existing Tabbit session can be reused when the CLI targets the same Tabbit environment, but it does not bypass a site login, CAPTCHA, permission gate, or account policy. In a local test fixture, a second CLI task reached a protected page after a short-lived test session was established. That verifies session reuse for the tested Tabbit environment, not an external Kimi integration.
Separate model, client, and browser routes
| Component | Role in this workflow | What this article does not establish |
|---|---|---|
| Kimi K3.1 Pro | Generates plans and tool-call requests | That the model can open your browser on its own |
| OpenCode or another local client | Sends prompts to the model and dispatches allowed tools | That every client accepts every Kimi provider's tool format |
| Tabbit CLI | Operates pages in a Tabbit Browser environment | Native Kimi integration or automatic /tabbit support in all clients |
| Kimi WebBridge | A separate Kimi browser-control route mentioned in community discussion | Compatibility with Tabbit CLI; it should not be treated as the same connector |
A Reddit user described using Kimi WebBridge with another coding agent and an invoice workflow. That is a report about WebBridge, not Kimi K3.1 Pro or Tabbit CLI. It is a useful reminder that browser connector, client, and model are separate choices, but it is not proof this setup works. Read the original discussion.
A second user reported that Kimi K3.1 Pro wrote a headless browser script to check a Three.js scene. That test checked rendering and console output. The invoice workflow below needs a different check: confirm the intended account and invoice ID on the live page. Read the Kimi test report.
When this setup fits
Use this arrangement when an authorized employee needs to cross-check records in pages they can already access, prepare a draft, and review it before a consequential action. It is a poor fit for bulk changes, payment approval, or unattended work. Keep the final save or approval in the human checkpoint unless your organization has a separately approved workflow.
If your first test fails, diagnose the layer: Does the client list the model? Does it produce a tool call? Can the local process run the Tabbit skill? Does Tabbit CLI reach the correct browser instance? Is the website already signed in? These checks narrow the failure without handing the agent broader permissions.
For setup details, see the Tabbit CLI complete guide, connect an agent to an existing browser, and the first browser-agent task checklist. Compare browser MCP and CLI if you are choosing the connector, or follow the Codex Tabbit CLI setup for that specific client.
For a focused overview of session behavior, see how browser agents reuse a login session. If the agent cannot run the connector, follow how to let an AI agent control your browser. For a review point inside a content system, compare the CMS draft workflow.
FAQ
Can Kimi K3.1 Pro use my existing browser login?
The model does not receive a browser session by itself. A local agent client must run a browser connector such as Tabbit CLI, and the target site must already be signed in within the Tabbit environment the CLI uses.
Does Kimi Code include Tabbit CLI?
This article does not verify a native Kimi Code and Tabbit CLI integration. Kimi Code is Moonshot's coding client; Tabbit CLI is a separate local browser tool. Use a client only after confirming it can load the Tabbit skill and execute its local commands.
Is Kimi WebBridge the same as Tabbit CLI?
No. Kimi WebBridge is a separate browser-control route. Community reports about WebBridge do not demonstrate Tabbit CLI compatibility or prove that Kimi K3.1 Pro was the model used.
Does an OpenRouter listing mean the endpoint will work in my client?
No. A provider listing shows a route is available in the directory, not that your account has access or that your client supports that provider's tool-call format. Test with a harmless, read-only task.
Can the agent submit an invoice or approve payment?
The example explicitly forbids edits and approval. Keep financial actions behind a human review checkpoint and follow your organization's access rules.
Can I keep working while the task runs?
Tabbit CLI uses task tabs in the target Tabbit environment. You can continue working in separate tabs, then return to inspect the named draft and its source values.