Machine Internet UAA

Point it at any URL. Get an MCP server any AI agent can call.

Live: https://machineinternet-production.up.railway.app

---

The problem

Most APIs on the internet are not accessible to AI agents. They speak HTTP, not MCP. They have no agent SDK, no tool definitions, and often no public documentation at all. Every new integration is a custom build.

Machine Internet UAA eliminates that. Give it a URL and it produces a working MCP endpoint in seconds — one that any agent, Claude, Cursor, or any MCP client can call against the real service with no additional code.

---

How it works

Step 1: Discovery. It tries to find an OpenAPI spec automatically by probing 20 standard paths. If it finds one, it parses it. If there is no spec, it launches a headless browser, observes the XHR traffic the page makes, and uses an LLM to infer the schema from what was captured. Every API on the internet is a target.

Step 2: Condensation. Raw API specs can have hundreds of endpoints. An LLM collapses them into 10-15 clean, agent-friendly tools with verb_noun names and descriptions that tell agents exactly when to use each one. A 400-endpoint CRM becomes get_customer, create_deal, log_interaction.

Step 3: Serving. The condensed schema is saved locally and served as a standard MCP SSE endpoint. Paste the URL into any MCP client. The agent makes real calls against the real service.

---

Quickstart

git clone https://github.com/chetanty/machine_internet
cd machine_internet
pip install -r requirements.txt
playwright install chromium
cp .env.example .env   # add GEMINI_API_KEY or OPENAI_API_KEY

Wrap any API with a public spec:

python discover.py --url https://api.github.com \
  --spec https://raw.githubusercontent.com/github/rest-api-description/main/descriptions/api.github.com/api.github.com.json \
  --tags repos,issues
python serve.py --schema schemas/github_v3_rest_api.json --port 8100

Add http://localhost:8100/mcp to your Claude MCP config. Done.

Wrap an API with no spec using traffic sniffing:

python discover.py --url https://hn.algolia.com --traffic
python serve.py --schema schemas/algolia_api_hacker_news.json --port 8101

Run the dashboard to manage everything from a browser:

python dashboard.py   # http://localhost:7000

---

What it can wrap

Service	Method	Tools generated
GitHub REST API	OpenAPI spec	14 tools: repos, issues, branches
httpbin.org	OpenAPI spec (auto-detected)	15 tools
Stripe, Petstore, any OpenAPI service	OpenAPI spec	up to 15 tools
PokéAPI	Traffic sniffing (no spec)	get_pokemon
HN Algolia search	Traffic sniffing (React SPA, API on different TLD)	search_articles, check_service_status
Open Library	Traffic sniffing (partial, SSR-limited)	search_books_authors, get_search_facets

Traffic sniffing works on SPAs that load data via XHR. Server-rendered pages yield fewer captures. The right target is any page that fetches data dynamically after load.

---

Claude MCP config

{
  "mcpServers": {
    "github": {
      "url": "http://localhost:8100/mcp"
    },
    "algolia": {
      "url": "http://localhost:8101/mcp"
    }
  }
}

---

Configuration

Copy .env.example to .env and fill in at least one key:

GEMINI_API_KEY=your-key        # free tier: 20 req/day on flash
OPENAI_API_KEY=sk-...          # no daily quota, gpt-4o-mini default

The AI client tries OpenAI first, then Gemini, then fallback Gemini models, then secondary Gemini keys. Add more keys to GEMINI_API_KEY_2 through GEMINI_API_KEY_5 to increase capacity.

---

Auth

Credentials are stored encrypted via Fernet in ~/.uaa/vault/. The server injects them per request as Bearer token, API key header, Basic auth, or OAuth2 client credentials flow.

python serve.py --schema schemas/github_v3_rest_api.json --set-creds

---

Deployment

Deployed on Railway via Docker. The live dashboard at https://machineinternet-production.up.railway.app runs discovery and wrapping in the cloud. MCP servers launched from the deployed instance are accessible within the container network.

To deploy your own instance:

railway up

---

Project structure

discover.py       discover any API and save a condensed schema
serve.py          serve a schema as an MCP endpoint
dashboard.py      web UI: wrap, start, stop services

src/
  discovery/      Path A (OpenAPI) and Path B (traffic sniffing)
  condensation/   LLM-based schema condensation and eval runner
  serving/        MCP server (pure ASGI SSE) and executor
  auth/           encrypted credential vault and request injector
  ai/             Gemini + OpenAI fallback client

evals/            ground truth schemas and scoring

For a full technical breakdown of the architecture, design decisions, eval results, and build log, see BUILD_REPORT.md.