Skip to content

BROWSER.md

Latest commit

 

History

History
89 lines (66 loc) · 2.57 KB

BROWSER.md

File metadata and controls

89 lines (66 loc) · 2.57 KB
Raw

Browser Module

The Browser module manages local Playwright browser automation for MatrixClaw. It is separate from web search: web research returns compact facts and sources, while browser tools can interact with rendered pages when a real browser is needed.

Open it from the terminal UI:

/modules browser

Provider

The current provider is Local Playwright:

{
  "modules": {
    "browser": {
      "enabled": true,
      "provider_id": "playwright",
      "provider_config": {
        "runtime_mode": "per_task"
      }
    }
  }
}

The module reports runtime installation state, managed Chromium installation state, runtime path, browser executable path, browser cache path, runtime mode, and runtime process state when a managed process is running.

Install And Repair

Use /modules browser -> Install/Repair to install the managed Playwright MCP runtime and its managed Chromium browser cache.

MatrixClaw treats the browser as installed only when the required Playwright Chromium revision contains a real executable. A stale or mismatched browser revision is shown as Repair Required, and running Install/Repair removes stale Chromium revisions before reinstalling the required one.

Run Modes

Local Playwright supports two runtime modes:

  • per_task: start an isolated browser runtime for the current job and stop it afterward. This is the default for lower idle memory.
  • always_running: keep a managed browser profile warm for lower startup latency.

The TUI exposes this as:

/modules browser -> Runtime Mode

MCP Relationship

Interactive browser actions are exposed to the assistant as MCP tools. When the Browser module is enabled and installed, the daemon injects a managed MCP server with ID browser. The MCP module reserves that ID so a user-defined MCP server does not conflict with the managed browser server.

Remote tools are registered with prefixed IDs such as:

mcp_browser_navigate
mcp_browser_click
mcp_browser_type
mcp_browser_screenshot
mcp_browser_wait

Exact names depend on the MCP server's tool names. Non-read-only browser tools use the normal MatrixClaw approval flow.

Web Research Relationship

web_research can use a browser fallback for pages that direct HTTP fetching cannot read. That fallback stores raw page artifacts and returns compact facts and source references to the assistant.

For interactive browser work such as clicking through flows, filling forms, waiting for dynamic content, or taking screenshots, use the MCP browser tools provided by this module or by another configured MCP browser server.