Web Access

What web-access does

The web-access MCP server exposes String AI's Web Access API through MCP-compatible clients. It gives an agent four operations: retrieve one URL as Markdown, search the web for structured results, send a POST/PUT/PATCH request with a body, and create or manage asynchronous site-crawl jobs.

Fetched pages are returned in a format intended for language-model consumption. According to the project documentation, pages affected by rate limits, geographic restrictions, or automated-traffic blocks are returned as Markdown rather than a block page. Proxy rotation, challenge handling, and JavaScript rendering are handled by String AI's service.

How it works

An MCP client starts the package as a child process and communicates with it over stdio using JSON-RPC. The process uses STRING_AI_API_KEY to call the String AI Web Access API over HTTPS, then passes the API response back to the client.

Sitemap operations use a quote-and-approval workflow. A submission returns a job ID, estimated cost, estimated page count, and an awaiting-approval status. Nothing is crawled or billed until the approval action is sent. The job can then be polled, read through paginated results, canceled while non-terminal, or listed with other recent account jobs.

Setup and configuration

Install and run the web-access MCP server with Node.js and the package published as @usestring/mcp:

env STRING_AI_API_KEY=your-key npx @usestring/mcp

For MCP client configuration, use npx with the -y flag and set STRING_AI_API_KEY in the server's environment. The README provides configurations for VS Code, Cursor, Windsurf, and Claude Desktop. The API key is required. The repository also documents global npm installation and running a locally built node build/index.js entry point.

The hosted deployment described by the project is available at https://mcp.usestring.ai/v1/mcp. Its four-tool interface differs from npm package version 1.2.1: that package routes writes through web_access_fetch using method and body, rather than exposing the separate request tool described for the hosted server.

Tools and capabilities

• web_access_fetch: Fetches one URL and returns clean Markdown. It is read-only in the hosted four-tool interface.
• web_access_search: Searches the web and returns structured results.
• web_access_request: Sends POST, PUT, or PATCH requests with a request body. This operation writes to the target URL.
• web_access_sitemap: Submits, approves, polls, reads, cancels, or lists crawl jobs. Submission supports page, depth, path-prefix, budget, and sitemap options; results are paginated.

Crawl limits include up to 10,000 pages and a maximum depth of 100. Result limits are also bounded, with a maximum page size of 5,000 for result retrieval.

Limitations and notes

The web-access MCP server requires a String AI API key, and sitemap crawling creates billed jobs. Approval is explicit, but already fetched pages remain billed if a job is canceled. Crawl results are durable after completion, while per-URL discoveredUrls data is available for approximately one hour.

A crawl can fail, be canceled, or exceed its token cap. The API may return insufficient-funds or partial-state conditions during approval; the documentation says to retry approval for the latter. The npm package and hosted endpoint do not expose exactly the same write-tool shape, so client configurations should match the deployment being used.

Getting started with this web-access MCP server

Always refer to the official documentation for the most accurate and up-to-date information.