DialMCP Connector
Give your AI agent a phone.
Open-source connector for DialMCP โ the hosted MCP server that lets AI agents place real phone calls from your own verified number.
What this is
DialMCP is a remote MCP server at https://mcp.dialmcp.com/mcp. Any client that speaks
Streamable HTTP with OAuth can connect to
it directly โ no install, no API keys. You do not need this package for those clients.
This repository exists for two reasons:
- A stdio bridge for clients that only support local servers. Some MCP clients still can't talk to
remote servers.
dialmcp-connector launches a local stdio server that proxies to the hosted endpoint and
handles the OAuth browser flow for you.
- The open, canonical documentation of the DialMCP integration โ tool schemas, client configs, the
server.json registry manifest, and the safety model โ in one auditable place.
The hosted calling service itself is closed-source. This connector is MIT-licensed and contains no
proprietary logic: it is a thin transport shim over a public endpoint.
Quick start
Direct connection (preferred)
If your client supports remote MCP servers, skip this package entirely:
URL: https://mcp.dialmcp.com/mcp
Transport: Streamable HTTP
Auth: OAuth 2.1 (sign in with your phone number)
Claude Code
claude mcp add --transport http dialmcp https://mcp.dialmcp.com/mcp
Codex โ ~/.codex/config.toml
[mcp_servers.dialmcp]
url = "https://mcp.dialmcp.com/mcp"
claude.ai / Claude Desktop
Settings โ Connectors โ Add custom connector โ paste https://mcp.dialmcp.com/mcp.
Gamut
Settings โ Connectors โ Add MCP server โ paste the URL (HTTP / OAuth) โ complete SMS verification.
On first connect a browser window opens for sign-in. Verify your phone number with an SMS code โ that
number becomes your caller ID.
Via this connector (stdio-only clients)
Or in an MCP client config:
{
"mcpServers": {
"dialmcp": {
"command": "npx",
"args": ["-y", "dialmcp-connector"]
}
}
}
See examples/ for ready-to-paste configs.
| Tool | Description |
|---|
place_call | Start an outbound phone call to a US or Canadian number with a stated objective. Returns immediately with a call ID โ the call runs asynchronously. |
get_call | Fetch the current state of a call: status, live progress, and once finished the transcript, recording link, and structured resolution. Poll this after place_call. |
end_call | Hang up an in-progress call immediately. |
list_calls | List recent calls for the authenticated user with their statuses and outcomes. |
The API is async by design: place_call returns instantly with a call ID and your agent polls
get_call, so a multi-minute call never trips a client's tool timeout.
When a call completes, get_call returns a structured resolution โ achieved, partially_achieved, or
not_achieved โ with a summary, any commitments the other party made, suggested follow-ups, a
turn-by-turn transcript, and a signed link to the audio recording.
During the call the service presses keys through automated phone trees, waits silently through hold music,
and detects voicemail โ hanging up rather than leaving a message.
Safety model
Every protection is enforced server-side, not as model instructions the agent could talk itself out of.
- Disclosure. Every call opens by stating it's an AI assistant, who it's calling for, and that the line
is recorded โ before the ask. Compliance is independently verified against each transcript and logged. The
agent never denies being an AI, and ends the call if the callee objects.
- Your real number. Calls present your own SMS-verified caller ID. Nothing is spoofed, and recipients
can call you back.
- Rate limits. 1 concurrent call ยท 3/hour ยท 10/day ยท max 2/day to the same destination ยท 6-minute
default and 10-minute hard cap, watchdog-enforced. Bulk calling is impossible by construction.
- Time and place. 8:00โ21:00 destination local time (TCPA-aligned), US and Canada only. Unknown time
zones are treated conservatively.
- Non-bypassable blocklist. Emergency services (911), premium-rate lines (900/976), and high-fraud
ranges can never be dialed.
- Permanent opt-outs. Do-not-call requests are confirmed on the call, caught by a second independent
transcript review, stored permanently, and enforced platform-wide across every user. Public self-serve
opt-out at https://mcp.dialmcp.com/opt-out โ no account needed.
- Objective screening. An independent reviewer screens each call's stated objective for harassment,
deception, or fraud before dialing. The agent states only facts supplied in the task.
- Prohibited uses. Telemarketing, cold outreach, robocalling, surveys, lead generation, debt
collection, and political outreach are barred by the Terms.
Full detail: https://dialmcp.com/safety.html
Requirements
- Node.js 18 or newer (only if you use the stdio bridge)
- A US or Canadian mobile number for SMS verification
- Users must be 18 or older
Links
Releasing
Every release publishes to two places: npm (dialmcp-connector) and the
official MCP Registry (com.dialmcp/dialmcp). Both are
automated by .github/workflows/release.yml.
Cutting a release
- Bump
version in both package.json and server.json โ npm run validate:manifest fails if they
disagree, and CI runs it on every push.
- Merge to
main.
- Run the Release workflow manually (Actions โ Release โ Run workflow,
dry_run left checked) and let
it go green. This packs the tarball and validates server.json against the live registry schema without
publishing anything.
- Publish a GitHub Release tagged
vX.Y.Z. The tag must match package.json or the workflow stops before
publishing.
Step 3 is worth the two minutes: registry versions are immutable and cannot be unpublished. A bad
version can only be superseded by a higher one or marked deprecated.
What the workflow does
| Job | Runs on | Steps |
|---|
publish | releases and manual dry runs | typecheck โ validate manifest โ build โ smoke test โ tag/version match โ npm publish --provenance โ mcp-publisher validate |
registry | releases only | wait for npm propagation โ mcp-publisher login dns โ mcp-publisher publish |
--provenance attaches a signed attestation linking the tarball back to the commit and workflow run that
built it, which is why the job needs id-token: write.
The registry publish is a separate job for two reasons. It has to wait for npm: the registry proves
ownership by fetching the freshly published version and reading its mcpName field, which 404s until npm
propagates (the job retries for ~2.5 minutes). And it holds the signing key, which is scoped to the whole
com.dialmcp namespace rather than this one package โ so it lives in a protected environment. Keeping that
environment off the publish job is deliberate: its branch policy only admits v* tags, and applying it
job-wide would block manual dry runs from main outright.
One-time setup
| What | Where | Why |
|---|
NPM_TOKEN | repository secret | automation token with publish rights on dialmcp-connector |
MCP_REGISTRY_PRIVATE_KEY | secret on the release environment | Ed25519 key, base64; authenticates the registry publish |
release environment | deployment branch policy โ tags โ v* | keeps the namespace key off untagged runs |
v=MCPv1; k=ed25519; p=<public key> | TXT record on the apex of dialmcp.com | how the registry verifies domain ownership |
The TXT record goes on the bare domain, SPF-style โ a _mcp. selector is not read and fails with an opaque
signature error. Signatures carry a ยฑ15s timestamp window, so the runner's clock must be accurate.
Contributing
Issues and pull requests are welcome for the connector, docs, and client configs. Bugs in the hosted
service itself are best sent to support@dialmcp.com. Please read SECURITY.md before
reporting anything security-related.
License
MIT ยฉ Datawizz Inc. See LICENSE.