[FR] Add MCP Support

[christianliebel]

I’ve noticed a growing adoption of the Model Context Protocol (MCP), which has become the de facto standard for agentic AI applications. MCP enables compatible clients to interact with external services, such as conducting web searches, obtaining driving instructions, and navigating websites, among many other functionalities.

Here’s a screenshot of GitHub Copilot in Visual Studio Code, showing tools provided by the Playwright MCP Server. By using natural language input in the Copilot chat box, you can then instruct Copilot to navigate to a certain URL, click a certain button, etc.

Given its very powerful and stateful nature, it may, if at all, only be feasible to implement full MCP support in the Prompt API for extensions. However, work is in progress for stateless communication that could possibly be triggered by websites.

In practice, this means adding additional, external tools to the LLM setup. Here’s an example from smolagents (Source). By using the base_tools, the agent is capable of performing a web search to find a remedy for hangover instead of using its world knowledge.

with ToolCollection.from_mcp(server_parameters, trust_remote_code=True) as tool_collection:
    agent = CodeAgent(tools=[*tool_collection.tools], add_base_tools=True)
    agent.run("Please find a remedy for hangover.")

Requires #7.

[zolkis]

I'd love we could support a subset of MCP APIs, since it would give a good surface to implement e.g. RAG use cases.

The big question is what MCP API do we expose to the client code: setting up tools, orchestration, guardrails etc, or what part (of guardrails for instance) is controlled by the browser.

[AdamSobieski]

I hope that you find interesting that the AI Agent Protocol Community Group recently launched. Their intended topics of exploration include, but are not limited to, security and privacy mechanisms: authentication, authorization, verifiable-credential-based trust, and end-to-end encryption. Some of these topics, e.g., a lack of authentication and authorization mechanisms, have been criticized with respect to the initial version of the MCP. See also the A2A protocol which addresses these concerns.

[anssiko]

I hope that you find interesting that the AI Agent Protocol Community Group recently launched.

That group is working on a white paper that provides an overview of the current agent protocols. The paper is good background reading to folks interested in the protocol landscape. That group's scope and focus is on protocols and browser integration is left to other groups.

In this group we're interested in how the applicable agent protocols can be leveraged by Web APIs. This issue explores Prompt API MCP support and issue #7 tool/function calling, a prerequisite for MCP support.