AI agent access

Pages offer their tools to AI agents in the browser through WebMCP, a proposed web standard that some browsers offer as a trial: an agent can fill in a tool and read the result, and you see it happen on the page.

Assistants without a browser can read the pages and link people to the tool that answers their question. Every page shows the same content to people and to machines.

For language models and agents, in English: an overview at /llms.txt, every tool's explanation and FAQ at /llms-full.txt, the catalogue at /catalog.json and the agent tool definitions at /agent-tools.json. Other languages have the same files under their language path, for example /fi/llms.txt.

A tool that is retired or replaced keeps a notice at its address, with a link to any successor, and can no longer be run. A tool that is temporarily unavailable shows a notice and is left out of the catalogue until it returns.

The agent guide provides endpoints, request examples and error codes in English.

Agents without a browser can connect to https://how.box/mcp using Streamable HTTP, without signing in. The server is stateless: it keeps no session between calls. The site's API stays private; /mcp is the one public interface for agents.

• Five catalog tools: catalog_search_utilities, catalog_list_facets, catalog_get_utility, catalog_describe_tool and catalog_run_tool. They find and describe tools and run read-only calculations, including ones that use our cached reference rates, such as currency conversions. No network checks or tools that need the browser. Every answer includes a link to the tool page.
• Limits: 60 requests per minute per client; each server instance accepts 20 requests per second and 8 calls at once. Each call has 5 seconds and a 64 KiB request-size limit. Requests over a limit are refused.
• Inputs and results are never logged or kept. Call logs record the method, catalog tool name, outcome, duration and protocol version; rate lookups also record the action, response status and duration.

Agents are treated exactly like people: nothing about them or their users is stored, and the same usage limits apply.

When you use our results, please link to the tool page you used, so people can check the answer and change the inputs.

On every page an agent can search the catalogue (by category, tag and country), list its categories, tags and countries, open a tool, or run any tool's calculation from the current page.

Most tools run entirely in your browser and have no limits. The few that use our server have the limits below, for people and agents alike. A visitor is recognised only by a code that changes every day; nothing about you is stored. A request over a limit, or one that arrives while the server is busy, is refused with the time after which you can try again.

“Check again now” asks for a new check instead of a recent saved answer. Between new checks of the same address there is a short wait, shown below; during it you get the most recent answer with its age, or a request to try again later.

Our robots.txt tells crawlers what they may read. It never changes what a page says: everyone gets the same pages. Classic search engines may always crawl the site.

Policy in use: open (robots.txt)