Skip to main content
The meetergo MCP server gives an AI agent real access to a meetergo account. It can check availability, book, reschedule and cancel meetings, manage contacts and routing forms, and set up the Mira website assistant end to end from nothing but a website address.
Available on every plan, including Free. All you need is a Personal Access Token from Integrations → API.
This is not the documentation search server. developer.meetergo.com/mcp searches these docs and cannot change anything. The server described here books meetings and writes to a real account. See Documentation MCP Server for the other one.

​
Connect

The hosted endpoint is:
The /mcp path is part of the endpoint. mcp.meetergo.com on its own is not it. Claude Desktop users can skip configuration entirely: download meetergo.mcpb from the latest release and open it. The token is collected in the install dialog and stored as a sensitive value.

​
Authentication

The simplest route, and the one that works everywhere. A PAT starts with rgo- and always acts as its owner — it cannot impersonate another user.Send it as a bearer credential:
Scope it when you create it. The capability groups (scheduling, CRM, Mira, forms, account) let you give an agent only what it needs.
The hosted endpoint publishes RFC 9728 protected-resource metadata, so an OAuth-capable client discovers the flow on its own:
Clients are pre-registered. mcp-claude is the public client and uses PKCE with no secret, so it is safe to configure directly.Dynamic client registration is deliberately not enabled: it is superseded by Client ID Metadata Documents in the authorization spec, and a directory client that registers on every connection would fill the realm with throwaway records.
A Platform API Key (ak_live:<uuid>:<secret>) names no user of its own, so it must say who it is acting as:
The server fails at startup if one is supplied without the other, rather than answering the first call and then breaking on the second.

​
Set up an entire account with one prompt

The server ships two guided flows as MCP prompts. Clients with prompt support surface them natively as meetergo: onboard and meetergo: weekly-review.
1

Orient

get_setup_status returns the launch checklist as the dashboard computes it, so the agent never rebuilds what already exists.
2

Analyse

propose_conversion_setup reads a sample of the website and proposes a persona, grounded instructions, qualification questions and the site’s own legal links. Nothing is stored.
3

Approval gate

The agent presents the proposal and stops. Nothing is written before the user explicitly approves.
4

Build

Meeting types, the knowledge base crawl, a real routing form for qualification, and the assistant configuration — with the website chat still switched off.
5

Prove

run_test_drive sends scripted visitors through the customer’s own assistant (a buyer who books, a lead who wants a callback, an adversary probing for invented promises) and returns pass/fail verdicts with full transcripts. Preview mode: no bookings are created and no email goes out.
6

Install and go live

The embed snippet, then verify_widget_install to confirm the loader and the account’s own key are actually on the page, and only then the switch to live.

​
What the tools cover

Scheduling

Availability, book, reschedule, cancel, list appointments, add guests, update notes

Meeting types

Create and edit, one-time booking links, booking page and personal page settings

Contacts

Create, update, search, bulk import, tags and account owners

Website assistant

Crawl a site into the knowledge base, write instructions, build qualification forms, test drive, verify install

Routing forms

Create, update and send forms; qualifiers and fallback rules

Webhooks

Register, update and remove endpoints
Every tool carries MCP annotations (readOnlyHint, destructiveHint), so a host can confirm with the user before anything irreversible happens. Those annotations are honest: run_test_drive is marked as a write because it persists its verdict and spends model budget, and reschedule_appointment, update_appointment_notes and update_contact are marked destructive because each one replaces something a person would miss.

​
Plan limits are answers, not walls

When an account’s allowance runs out, tools return the limit structurally — which feature, what the next tier unlocks, and the upgrade link — instead of an opaque error. get_me reports the tier and caps up front so an agent can warn before starting something the plan cannot finish. The connection itself is never gated by plan, and the tool descriptions instruct the agent not to raise upgrades unless a requested action is actually blocked.

​
Security model

A Personal Access Token cannot impersonate another user; the API rejects the impersonation header outright. Capability groups let you hand an agent the least it needs.
The hosted server never forwards a credential handed to it by an OAuth client. An access token is validated against our Keycloak realm — issuer, audience, signature, expiry — and exchanged for a separate token before any API call, as the MCP authorization spec requires.
The one tool that fetches a caller-supplied URL (verify_widget_install) resolves the hostname first and refuses loopback, link-local, and private addresses, re-checking every redirect hop. The server cannot be used to reach a network it can see and you cannot.
Results that carry third-party text — website-visitor questions, crawled pages, test-drive transcripts — are wrapped in an explicit untrusted-content marker telling the agent to treat them as data and never as instructions.
The onboarding flow writes nothing before explicit approval, and every Mira settings change returns a snapshot that restore_mira_settings can put back.

​
Troubleshooting

The token never reached the server. In a config file it belongs inside the env block; against the hosted endpoint, check the header reads Authorization: Bearer rgo-….
The token is wrong, was revoked, or was pasted with whitespace. Create a fresh one at Integrations → API. The WWW-Authenticate header on the response names the reason.
The token was scoped without that capability group. Create a new token that includes it.
It runs several real conversations, so it is capped at a few runs per hour per company. The saved configuration is unaffected.
If the site is public and renders normally, this is almost always our model lane being slow rather than anything wrong with the website. Retry in a moment.

​
Source

Open source at github.com/meetergo/meetergo-mcp-server, published as @meetergo/mcp-server.