One install for the assistant, one signup for the business. The assistant calls typed verbs against real backends. The business sees the record the same way a human submission would land. No special agent inbox to maintain.
They land on a human marketing page. They guess at buttons. They click around hoping a form is what the user asked for. They sometimes succeed; mostly they apologise and ask the user to do it themselves. The interface is the wrong shape for the worker.
A typed contract per capability. Declared invariants. A backend that honours them. Errors that can be acted on. Idempotency where it matters. Latency that fits a sync conversation. No marketing copy in the response.
A small, shared vocabulary of verbs across CRM, scheduling, lead intake, and project work. Each verb has a spec page with input and output schemas, declared invariants, and a public score from real backend invocations every release. One MCP install reaches every business inside.
Each one is a typed contract: what the assistant gives, what it gets back, what cannot change while inside. The list extends. New verbs go through the spec, get a conformance suite, then ship. Browse the live registry at /capabilities.
The universal config runs through a local bridge that every desktop assistant understands. The direct config is for newer assistants that already speak the modern transport.
{
"mcpServers": {
"whatcanido": {
"command": "npx",
"args": [
"-y",
"mcp-remote",
"https://whatcanido.dev/api/mcp"
]
}
}
}{
"mcpServers": {
"whatcanido": {
"url": "https://whatcanido.dev/api/mcp"
}
}
}Per-assistant guides: Claude Desktop · Claude.ai · ChatGPT · Cursor · VS Code Copilot.
Each is a real business with services, hours, calendar, an inbox. Pick one and walk through it the way the customer would. The assistant calls find_providers → get_provider_actions → submit_action.
Eight-week career resets for senior engineers thinking about the next step.
Honest roof work. Free estimates within 24 hours.
Stomatologická klinika přijímající nové pacienty
The agent web is about to fill with "trust me" claims. Ours is the opposite. Every verb is exercised against its provider's real backend, every release. Test by test. Pass or fail. Published. When a provider's score moves, the assistant sees it and routes around them. The pressure goes the right way.
Whatcanido is where a small business becomes reachable by AI assistants the same way it is reachable by humans on the web. One install for the assistant, one signup for the business.
Each SaaS vendor that ships its own MCP server teaches the assistant a new vocabulary, with new field names. Multiply that by every vendor and you get a standard nobody wrote. Whatcanido sits one level above and gives the assistant a shared vocabulary that means what it says.
Open Claude Desktop config (Settings → Developer → Edit Config), add this to mcpServers: { "whatcanido": { "command": "npx", "args": ["-y", "mcp-remote", "https://whatcanido.dev/api/mcp"] } }. Save, restart, done.
Settings → Connectors → Add custom connector → https://whatcanido.dev/api/mcp. No OAuth required for the public surface.
When a customer asks an AI to do something at your business (book, ask, quote, look up an invoice), the assistant reaches you the same way a human form submission would. Your existing flows fire. You don't operate a separate agent inbox.
The spec, the protocol, and the conformance test runner are MIT-licensed. The SaaS products that run on top are commercial.
Every capability is exercised against the real backend on a schedule. Every result is published on the capability detail page. If a provider stops behaving, the score moves immediately and the assistant can route around them.