An agent can analyze a request, choose the right next step, and call every available API—then stall when the final action exists only in a web interface.
AI agents are increasingly capable of reasoning, planning, and working with APIs. But many real business processes still depend on websites: business portals, admin consoles, support tools, forms, dashboards, and applications without an API for every action. That last mile can stop an otherwise capable agent.
The Playwright Workspaces remote Model Context Protocol (MCP) server, now in preview closes this gap. It gives MCP-compatible agents a managed remote browser and a comprehensive set of browser automation tools optimized for AI agents—without requiring teams to install Playwright, package browser binaries, or operate browser infrastructure in the agent environment.
The remote MCP server uses Streamable HTTP—a transport protocol that allows real-time, bidirectional communication between your agent and the browser session—and is delivered as a Playwright Workspaces feature. Agents can create a browser session, navigate and inspect pages, interact with elements, upload files, review console and network activity, capture screenshots, and close the session when the workflow is complete. We are exposing definitive set of 22 tools which can help you complete browser flows successfully with agent interactions.
MCP config with GitHub Copilot CLI:
MCPTool(
server_label="playwright-browser-automation",
server_url=mcp_server_url,
headers={"x-api-key": playwright_service_access_token},
require_approval="never",
)
Configure Tool approval Foundry:
Why this release matters
Enterprise workflows rarely live in one API.
The reality is:
- Critical tasks often end in a web interface.
- Multi-step workflows require both reasoning and interaction.
- Website structure and timing can change between runs.
- Agent environments might not allow browser packages or local processes.
- Teams need visibility and control when automation reaches an unexpected state.
The Playwright Workspaces remote MCP server closes this gap by exposing browser automation through an open protocol. Your agent uses standard MCP tools while Playwright Workspaces provides the remote browser infrastructure.
This makes it easier to add browser execution to an existing agentic workflow and keep the browser lifecycle explicit, observable, and bounded.
What's new
- A hosted remote MCP endpoint : Connect an MCP-compatible client to a workspace-scoped HTTPS endpoint (a URL unique to your Playwright Workspace that handles all MCP communication). The service is hosted and managed, so there is no local MCP server process to install or maintain.
- Browser tools designed for agentic workflows : The preview exposes 22 tools across the browser lifecycle. They cover navigation, accessibility-based page inspection, interaction, tabs, file upload, screenshots, browser diagnostics. See the product documentation for the complete tool and parameter reference. Accessibility snapshots give agents a structured way to understand a page and select action targets. For focused tasks, browser_find can locate relevant snapshot nodes without returning the entire page representation.
- An observe-act-verify automation loop : Reliable browser agents shouldn't execute a long sequence of clicks without checking the result. The tool set supports a practical loop which helps agents recover when a page rerenders, a selector becomes stale, a dialog opens, or an action completes differently than expected.
Step Action Tool Example 1. Observe Understand the current page state browser_snapshot or browser_find 2. Act Execute on a snapshot reference or selector browser_click, browser_type 3. Wait Pause for expected page condition browser_wait 4. Verify Confirm resulting state before continuing browser_snapshot 5. Close End the browser session when complete browser_close
- Support for Microsoft Foundry, GitHub Copilot CLI and other agent builder surfaces : You can connect the remote server to various agent builder surfaces and use cloud browsers for capabilities including session isolation and private website support. VNet injection lets agents access internal websites that aren't exposed to the public internet—useful for enterprise admin consoles and internal tools.
- Visibility using LiveView when you need investigation: A created session can return a Live View URL when Live View is available. This gives developers and operators a way to observe browser activity while evaluating or troubleshooting an agentic flow. The server also exposes bounded console-message, network-request, snapshot, and screenshot results. These signals help identify navigation failures, page errors, timing issues, and unexpected application state.
Choose the right Playwright experience
Playwright supports several development and automation models. Choose the experience that matches your workload:
|
What you need |
Recommended experience |
|
Give an MCP-compatible agent a hosted browser through a remote endpoint | |
|
Add browser automation to Microsoft Foundry through its integrated toolbox experience | |
|
Bring your remote CDP browsers with preferred choice of framework | |
|
Run an existing Playwright test suite at scale across managed cloud browsers |
The Remote MCP Server is designed for agent-driven, tool-based browser workflows. It complements rather than replaces the Playwright SDK, OSS playwright-cli, or the integrated Browser Automation Tool in Microsoft Foundry.
Get Started
- Build your endpoint URL: Take your Playwright Workspaces API endpoint, replace .api with .mcp and append /mcp to create the remote MCP server URL. Example: https://{region}.mcp.playwright.microsoft.com/playwrightworkspaces/{workspaceId}/mcp
- Configure authentication: Add the x-api-key header with your Playwright Workspaces access token. Example : set: x-api-key: <your-access-token>
- Connect your agent to MCP
- TIP : Exercise the Tools approval flow as present in Foundry, Copilot CLI– easier to get started
What this unlocks
The Remote MCP Server enables agents to complete workflows across APIs and websites, combine automation with human approval, inspect dynamic applications, and collect page, console, network, and screenshot evidence. Instead of stopping when a process reaches a browser, an agent can take a controlled action, verify the result, and continue toward the outcome.
Next Steps
- Try it now: Follow the MCP Quickstart to connect your first agent in under 10 minutes
- Explore the tools: Review the complete tool reference to see all 22 browser automation capabilities
- See it in action: Check out our sample agents repository for working examples with Foundry and GitHub Copilot CLI
- Get help: Join the conversation on Discord or submit feedback via our feedback channel