The builder door

The system of record your agent runs a local business on.

Contacts, jobs, appointments, conversations, consent, and the runs that answer, book, remind and brief, behind one MCP tool surface. Tools propose; the runtime executes. The rules are in the tools, not the prompt.

Connect

api.generaldispatch.ai

MCP endpoint (streamable HTTP):  https://api.generaldispatch.ai/mcp
Authorization:                   Bearer <api_key>
Instructions for your agent:     https://api.generaldispatch.ai/agent.md

An agent can open an account itself: POST https://api.generaldispatch.ai/signup returns the key. MCP clients that authenticate people use OAuth with PKCE; discovery is at /.well-known/oauth-protected-resource.

What the tools enforce, so your prompt does not have to

Every rich read returns a ctx token; every write that depends on it takes it back. A stale draft is denied at send time, which is the right answer.
send_message returns queued, needs_approval or denied, never sent. Delivery is re-checked against consent, takeover and your versions under locks.
Consent is recorded by the ingest and by humans. An agent can revoke it and can never grant it.
Approvals, confirmations and takeover are human decisions; approve cannot be granted to a key or a client.
Skills are text files on top of primitives: skill://inbound-triage-and-booking, skill://estimate-followup, skill://morning-brief, skill://onboarding.

Who it is for

Owners who already run their business from Claude or ChatGPT, and agencies who run many businesses from one agent. The hosted agent for owners without one uses exactly the same tools; there is nothing it can do that yours cannot.