One MCP server. Any call provider. Fully local, fully hosted, or bring your own frontier model.

CallMCP is a single Model Context Protocol server that defines one tool contract — 14 tools, identical schemas, identical error shapes — for outbound/inbound telephony: calls, SMS, recordings, transcripts, and phone-number lifecycle. Point it at a hosted backend, a fully local backend, or a bring-your-own-key composed backend, and your agent code doesn't change. What changes is which tools are present (via capability-gated dynamic discovery), never the shape of a tool that's there.

The normative reference is SPEC.md. This README is the pitch; the spec is the contract.

npx -y @callmcp/server

Before connecting a real provider, run npx -y @callmcp/server doctor. For a safe local walkthrough, run npx -y @callmcp/server --sandbox. The server fails closed when no driver is configured; the mock driver is only available through explicit sandbox configuration.

See examples/ for ready-to-paste Claude Desktop / Claude Code configs for each of the three legs below.

Three structural claims

This isn't "a Twilio wrapper with a marketing page." Three things are actually built into the contract, not bolted on:

1. Provider-neutral, multi-backend contract

Every driver — hosted, local, or BYOK — implements the same 14 tools defined in SPEC.md. Driver-specific behavior lives exclusively inside a namespaced options.<driver_id> passthrough object. There is no make_call_twilio or make_call_kaicalls. A tool either exists for your configured driver (and works, fully, per spec) or it's absent from tools/list — never present-but-broken. Swapping your call backend is a config change, not a rewrite.

2. x402-funded autonomous provisioning to production

search_numbers, buy_number, configure_number, and list_numbers are never human-approval-gated — they don't contact a third party, they contact your own provider account. That includes settling cost via x402 machine payments when a driver requires prepayment (INSUFFICIENT_FUNDS embeds a full x402 challenge an agent can settle and retry). An agent can search for a number, pay for it, configure it, and go live — autonomously, in production, with real money — without a human clicking anything. This is a deliberate contrast with vendor MCPs whose autonomous/self-serve paths are trial- or sandbox-scoped only.

3. Human-in-the-loop outbound approval as designed consent architecture

make_call and send_sms — anything that contacts a phone number that isn't the calling agent's own infrastructure — structurally require a valid approval_id or a standing allowlist match before they act. This isn't a policy suggestion layered on top; it's enforced in the server core, so no driver implementation can bypass it. Primary path is MCP elicitation; the fallback for non-elicitation clients is a single-use out-of-band approval URL — the gate is designed to never silently block forever and never silently open. Read as a TCPA-consent posture: the contract makes "who approved contacting this number, and when" a first-class, auditable object (request_call_approval / list_approvals), not an assumption baked into application code you have to get right yourself.

Related MCP server: Straight Connect

CallMCP vs. single-vendor MCPs vs. DIY

Every "(unverified, recheck before publishing)" cell is a placeholder, not a claim — we do not put a number or a fact about a competitor's product in this table without a citation behind it. If you're reading this before that recheck has happened, treat those cells as unknown, not as "no."

Who builds this, and what's actually free

CallMCP is built and maintained by CallMCP / KaiCalls. Full transparency, because hiding this costs more credibility than stating it:

• This OSS server is not a loss-leader funnel with the real functionality locked behind a paywall. It routes to any configured driver — hosted, local, or BYOK — with no markup and no artificial capability gating. A self-hosted Dograh driver or a BYOK Twilio+LLM driver gets the exact same 14-tool contract as the KaiCalls driver.

• KaiCalls is the hosted default — the path of least resistance if you want a call backend running in minutes with nothing to self-host. It is not the only supported path, and the spec and conformance suite exist precisely so that claim stays true as the driver ecosystem grows.

• The driver interface is public and the conformance suite runs in CI. Any driver — including ones we didn't write — can claim conformance, and community PRs are held to the same bar KaiCalls' own driver is held to (see SPEC.md §6).

If any of the above stops being true, that's a bug in this README, not a hidden asterisk — open an issue.

Driver capability matrix

Capability is expressed by presence, not by runtime errors (see SPEC.md §2.2). This table mirrors the spec's degradation appendix (SPEC.md §7) — the honest state of the wider backend landscape at spec-writing time (2026-07-09), not a ceiling on what CallMCP itself can do.

Launch drivers (this repo)

Known gaps across the wider backend landscape (context for driver authors)

A driver claiming a capability it can't demonstrate against the conformance suite fails CI (see SPEC.md §6.2) — this table is a snapshot, not a promise about backends CallMCP doesn't control, and any driver author whose backend's real capability differs from this snapshot updates their own manifest, not this README.

Repo shape

callmcp/callmcp
├── SPEC.md                   canonical tool contract (start here)
├── packages/
│   ├── server/                @callmcp/server — MCP core: transports, driver
│   │                          registry, config resolution, approval state
│   │                          machine, elicitation + fallback-URL flow,
│   │                          dynamic tools/list
│   ├── driver-interface/      @callmcp/driver-interface — TS interface,
│   │                          capability manifest types, conformance test harness
│   ├── driver-kaicalls/       hosted default driver
│   ├── driver-dograh/         local/self-hosted driver
│   └── driver-byok/           bring-your-own transport + bring-your-own LLM
├── examples/                  Claude Desktop / Claude Code config blocks per leg
├── server.json                official MCP Registry manifest — published as
│                              ai.callmcp/server (registry.modelcontextprotocol.io)
├── smithery.yaml               Smithery container deployment config
└── Dockerfile                  self-host / Smithery container build

Packages

Contributing a driver

Implement the interface in @callmcp/driver-interface, ship a callmcp.manifest.json per SPEC.md §6.1, and run the conformance suite. A capability your manifest doesn't claim true simply results in that tool being absent from tools/list — that's the whole mechanism, not a workaround.

License

MIT for this repository's code. SPEC.md is released under Apache-2.0 as documented in its own header, specifically so anyone can implement a conformant driver without asking permission.