Capture

In development

Everything on this page is a reserved surface that is landing incrementally. Tool names and argument shapes may change until the corresponding wave ships. Do not depend on these responses in production yet. When a tool ships, this page loses its “in development” marker for that section.

Capture covers three related surfaces:

• Accessibility snapshots and semantic actions (browser_click, browser_type, browser_select).
• Screenshots of the full viewport or a specific element.
• Network capture as a HAR archive, gated by the credential firewall.

Accessibility surface (in development)

The accessibility snapshot produces a structured, ref-based representation of the page suitable for LLMs. Every actionable node comes back with a stable @ref that the semantic action tools consume.

browser_accessibility_snapshot (in development)

Return a compact accessibility tree of the current tab. Non-interactive nodes are elided by default. Every actionable node includes a @ref.

Planned shape:

{
  "url": "https://example.com/signup",
  "tree": [
    { "ref": "@ref_a1", "role": "heading", "level": 1, "name": "Create account" },
    { "ref": "@ref_a2", "role": "textbox", "name": "Email",    "required": true },
    { "ref": "@ref_a3", "role": "textbox", "name": "Password", "required": true, "sensitive": true },
    { "ref": "@ref_a4", "role": "button",  "name": "Sign up" }
  ]
}

Nodes marked sensitive: true are still enumerated so the agent can find them, but their value is redacted anywhere it would otherwise appear.

browser_click (in development)

Click an actionable node.

Planned parameters: tab_id, ref.

browser_type (in development)

Type text into a text input.

Planned parameters: tab_id, ref, text, and an optional submit boolean that presses Enter after typing.

browser_select (in development)

Choose an option in a <select> element.

Planned parameters: tab_id, ref, value.

The action tools all take a @ref from a prior browser_accessibility_snapshot (or page_query_selector). Freeform coordinate clicks are not supported.

Screenshots (in development)

Two shapes, matching the two things agents ask for.

browser_screenshot_full (in development)

Screenshot the full viewport of a tab.

Planned parameters: tab_id, optional format (png default, jpeg, webp), optional quality for lossy formats.

Response is a base64-encoded image plus a mime_type. Suitable for direct handoff to vision-capable models.

browser_screenshot_element (in development)

Screenshot a specific element referenced by @ref.

Planned parameters: tab_id, ref, and the same format options as browser_screenshot_full.

Network capture (in development)

The network surface produces standard HAR archives. Because HAR bodies and headers routinely leak secrets, this surface is behind the credential firewall from the first byte.

Firewall-gated

The HAR export runs a header pass, a URL pass, and a body pass through the firewall before returning. Authorization, Cookie, and Set-Cookie are unconditionally redacted. Request bodies detected as JSON with fields matching the configured sensitive-key list are redacted too.

browser_network_start_capture (in development)

Begin recording network activity for a tab.

Planned parameters: tab_id, optional include_response_bodies (default false), optional max_entries.

browser_network_get_har (in development)

Return the current HAR archive.

Planned response:

{
  "har": {
    "log": {
      "version": "1.2",
      "creator": { "name": "Maho", "version": "0.1.0" },
      "entries": []
    }
  },
  "entry_count": 42,
  "truncated": false
}

browser_network_stop_capture (in development)

Stop recording and discard the buffered HAR.

Availability

The Agent Protocol’s tool list is discoverable at runtime with the standard MCP tools/list method. If a tool on this page appears in that list, it is shipped. If it does not, treat it as “in development” and check back after the corresponding release.

For the currently shipped tool surface see:

• Tabs
• Navigation
• Content