String AI Web Access MCP Server is the official Model Context Protocol (MCP) server for the String AI Web Access API, used to let LLMs access web content via the MCP protocol.
Purpose This server provides web access capabilities for MCP-compatible clients: search the web, fetch any URL or send requests to it, and map site URLs, returning results as clean Markdown suitable for LLMs. When web content is needed, the LLM can call web_access_fetch, web_access_search, or web_access_sitemap.
Main Tools
- web_access_fetch, web_access_search, and web_access_product_help are read-only operations.
- web_access_request is a write operation; web_access_sitemap creates a billable crawl task.
- Hosted server https://mcp.usestring.ai/v1/mcp provides six tools. The npm package provides five of them: it includes product help and failure reporting, but not web_access_request; write operations are performed via web_access_fetch's method and body parameters.
- web_access_sitemap: sitemap crawl task, an asynchronous two-phase process—quote, approve, poll, read; it will not crawl or bill before the quote is explicitly approved.
- web_access_report: failure diagnostics, optional and best-effort. You should first continue with useful recovery actions; if the report still has value and is permitted, send at most one report per task for each distinct failure, rather than sending one on every retry. Report exceptions, timeouts, tool errors, or unusable output; exclude usable source statuses, valid negative results, empty 204 responses, running sitemap tasks, and cancellation operations. Each report attempt has a 2-second deadline. Before calling, remove the Authorization header, API keys, Cookies, session tokens, personal data, and irrelevant conversation content; the server will again redact common credential forms. Never report web_access_report failures, and do not repeatedly call a failing tool just to collect report context. Reports are authenticated using the configured API key, but do not consume Web Access credits. If reporting is unavailable, unauthorized, rate-limited, or fails, stop reporting for that task.
Usage Installation: global installation or building from source is supported. Configure environment variables, then configure in the client.
- VS Code: press Ctrl+Shift+P → Preferences: Open User Settings (JSON) and add the configuration; or add a .vscode/mcp.json file to share the configuration with the team.
- Windsurf: open Settings → Features → MCP Servers → + Add new global MCP server and paste the configuration; or add it to ~/.codeium/windsurf/model_config.json.
- Claude Desktop: add it to claude_desktop_config.json. Testing: use MCP Inspector to interactively test the server in the browser—open http://127.0.0.1:6274，通过 stdio connection and call each tool in the UI. The IDE will start the server as a child process and communicate via stdio.
How It Works When the LLM determines that web content is needed, it calls web_access_fetch, web_access_search, or web_access_sitemap. Each failed call is reported exactly once via web_access_report, even if a subsequent recovery succeeds. The server uses the API key from the environment to forward requests to String AI's Web Access API and returns the results to the LLM.
Use Cases and Limitations Suitable for any MCP-compatible client (VS Code, Cursor, Windsurf, Claude Desktop, etc.), providing LLMs with web search, fetching, site mapping, and content-to-Markdown capabilities. Limitations include: web_access_sitemap creates billable tasks and requires approval of a quote first; web_access_report has count and time limits; the npm package and the hosted server have different tool sets; after a report fails, reporting should stop. The API key must be obtained at usestring.ai. Report security vulnerabilities privately according to SECURITY.md.
