Kimi Browser Extension Turns Recorded Actions Into Reusable Skills for Agents

In September 2026 Moonshot renamed its WebBridge product to Kimi Browser Extension and expanded its capabilities. The extension now works in both Chrome and Edge, allowing a signed‑in user to open pages, click buttons, fill forms and extract information directly from the browser.

The rollout was accompanied by a series of webinars and a detailed migration guide, ensuring that existing WebBridge users could transition without losing any of their previously recorded automations.

Live Chat Sidebar and Local‑Agent Control

A central feature of the new extension is a sidebar that hosts a live chat with Kimi. While users type queries, Kimi can act on the current page, performing the recorded actions in real time. This design keeps the interaction local to the browser, preserving privacy and reducing latency compared to cloud‑only solutions.

The sidebar also displays real‑time logs of each action, giving users immediate feedback on what the AI is attempting and allowing them to intervene if something goes awry.

From Recorded Clicks to Reusable Skills

The extension can record any sequence of actions a user performs on a website – navigation, button clicks, form entries, and data extraction – and then package that sequence as a reusable Skill. Once saved, the Skill can be invoked by any compatible local agent, effectively turning a manual workflow into a command‑line‑like interface for an AI assistant.

Each Skill is stored with a timestamp, a version number and a metadata block that describes the target domain, required permissions and any optional parameters that the agent may supply at runtime.

Kimi states that frequently used sites can be broken down into these command‑line‑style interfaces, making it possible for agents to execute complex browser tasks without human supervision each time. The company notes that this approach is especially useful for repeatable work such as reporting, form filling and quantitative back‑testing.

Support for Multiple Local Agents

• Kimi Work
• Claude Code
• Codex
• Cursor
• Kimi Code
• Hermes Claw

Kimi’s documentation lists six local agents that can consume the Skills generated by the extension: Kimi Work, Claude Code, Codex, Cursor, Kimi Code and Hermes Claw. Each agent can call a Skill via a simple API, allowing developers to embed browser automation inside larger AI‑driven pipelines.

The older remote‑control path that routed commands through a cloud server remains available for local agents that prefer that architecture. This dual‑path design ensures backward compatibility for existing integrations while encouraging migration to the more responsive local‑agent model.

Limitations and Best Practices

Kimi warns that page redesigns, dynamic content loading and JavaScript‑heavy interfaces can still disrupt automation. When a complex flow fails, the extension recommends inspecting the screenshots it automatically captures during the run to pinpoint where the recorded steps diverged from the live page.

The company also advises users to keep the browser window at a stable size and avoid pop‑ups that change the DOM structure during recording. These precautions help maintain the fidelity of the recorded Skill and reduce the need for manual adjustments later.

A successful agent session – where Kimi completes a task without errors – can be saved as a Skill for later reuse. The saved Skill includes the exact sequence of actions, any extracted data fields, and optional error‑handling logic that the agent injected during the run.

What this means for English‑speaking organisations is a concrete reduction in the time spent building and maintaining browser‑based automation. Teams can now record a workflow once, store it as a Skill, and let any of the supported local agents run it on demand, turning repetitive web tasks into a single line of code.

The result is faster onboarding of automation, lower reliance on custom scripting, and a clearer audit trail for compliance, since each Skill is versioned and tied to a specific recorded session.

In the Chinese edition of L'actu IA, the article appears with the same technical details but includes a brief note on how regional data‑privacy regulations influence the choice between local‑agent and cloud‑based execution paths.