Unlocking Hidden Site APIs for Small Business Automation
Machine Internet lets small businesses tap into the data behind any website and feed it directly into their workflows.

Instead of manual copy‑paste or fragile scrapers, you can expose a site’s hidden API as an MCP tool that works with Claude, Cursor, or any AI agent. The process costs only a few cents of LLM usage and can be run locally to avoid rate limits.
What Are Hidden APIs?
Many sites that small businesses rely on do not publish a public API. Their front‑end pages, however, make JSON requests behind the scenes to fetch data such as product lists, pricing, or news feeds. These undocumented endpoints are often called “hidden APIs” because they are not exposed for programmatic use. Examples include the Y Combinator company directory, Spotify podcast charts, Substack post searches, VS Code Marketplace listings, and bundlephobia package analyses. The data is real and up‑to‑date, but it is only accessible through the browser UI or custom scripts.
How Machine Internet Turns Them Into Tools
Machine Internet automates the discovery and exposure of those hidden APIs. It first probes about 20 standard locations ( /openapi.json , /swagger.json , … ) for an OpenAPI spec. If a spec exists, it is parsed directly. When no spec is found, the tool launches a headless Chromium instance, loads the target page, scrolls and clicks through it, and records every same‑domain JSON request. Analytics and ad trackers are filtered out. An LLM then reads the captured requests and responses, infers each endpoint’s path, parameters, and data types, and condenses the endpoints into at most 15 verb_noun tools. Each tool is described so an AI agent knows when to use it. The result is a small JSON file that can be served as a standard MCP server (stdio by default, or SSE with a flag). Clients such as Claude Desktop, Cursor, or any MCP‑compatible agent can call the tools directly, hitting the same endpoints the site uses.
Why It Matters for Small Business Owners
Traditional automation required developers to write custom scrapers or negotiate with vendors for an official API. Both approaches add cost, maintenance overhead, and risk of breaking when the site changes. Machine Internet removes those barriers: a boutique bakery can pull daily ingredient prices from a supplier’s site that offers no public API, feeding the data straight into its ordering system. The same workflow can be used to sync competitor pricing, monitor job postings, or gather market intelligence without hiring a programmer. Because the tool runs on a free LLM tier (Gemini) for roughly 20 wraps per day, the expense is minimal. For heavier use, the server can be run locally, bypassing cloud rate limits and ensuring reliability.
Getting Started Without Breaking the Bank
The only prerequisites are uv (a fast Python package manager) and an LLM API key. Gemini’s free tier provides a daily quota that covers about 20 wraps; if the quota is exhausted, Machine Internet falls back to Gemini Flash‑Lite automatically. Adding an Anthropic or OpenAI key is optional and only needed when the free tier is insufficient. The headless browser downloads once (about 120 MB) and is reused for every wrap. Wrapping a site takes 1‑3 minutes and costs a few cents of LLM usage on paid tiers, or nothing on Gemini’s free plan. Wrapped servers are saved to ~/.machine-internet/schemas/ for reuse. For sites that rate‑limit cloud requests, running `machine-internet serve` locally routes traffic through your own connection, avoiding throttling. The tool is also available as a hosted remote MCP endpoint at machine-internet.vercel.app, so no local installation is required for quick testing.
Takeaways
- Hidden APIs give access to real‑time data that most sites never expose, and Machine Internet automates the process of turning those calls into ready‑to‑use tools.
- The solution costs only a few cents per wrap on paid LLM tiers and can be used for free on Gemini’s daily quota, making it affordable for any small business.
- Running the MCP server locally avoids rate limits and ensures reliable automation for high‑volume tasks.
- The tool works with any AI agent that supports MCP, including Claude Desktop and Cursor, and can be integrated into existing workflows without custom code.
Where would this work in your business?
Take the two-minute AI Opportunity Snapshot and we will point to where automation is most likely to pay off for you.