GoHighLevel MCP Server: What It Is and How It Works (2026)

The useful question in 2026 is no longer whether the vendor ships an MCP server. It is how much of the product that server can reach, and under what permissions. That matters for agencies evaluating a switch, because an MCP server is the difference between an AI layer that can only read what you paste into it and one that can act inside the system of record. If you are weighing platforms, our Conduyt vs GoHighLevel comparison covers the wider feature and pricing picture. If your team works in Claude, the Claude Desktop CRM connection lets the assistant read and update records through MCP.

1. Pick the endpoint for your client. Claude uses https://services.leadconnectorhq.com/mcp/anthropic/v2; ChatGPT, Codex and compatible OpenAI clients use https://services.leadconnectorhq.com/mcp/openai/v2/; Muse connects through its official LeadConnector app at https://services.leadconnectorhq.com/mcp/muse/v2/.
2. Add it as a custom connector. The developer portal gives steps for Claude.ai, Claude Code, Claude Cowork, ChatGPT developer mode and Codex.
3. Complete the LeadConnector OAuth flow: sign in, choose the sub-account or sub-accounts, and approve scopes. Access can be reviewed or revoked from your LeadConnector account.
4. Confirm with a simple request, such as finding the last five contacts added.

• list_locations: list the sub-accounts the connection can use.
• search_operations: find operations by intent, including record searches and create, update or delete actions.
• describe_operation: inspect an operation’s inputs, required scopes and safety information.
• execute_operation: run an operation, subject to scopes, permissions and safety checks.

The documented working pattern is search_operations, then describe_operation, then execute_operation. HighLevel lists coverage across contacts, conversations and messages, opportunities and pipelines, calendars and appointments, payments, products and store, invoices and estimates, social planner, blogs, emails, forms and surveys, funnels, workflows, Voice AI, phone system, knowledge base, media, custom objects, users, and ads and affiliates.

How large is the catalog behind those tools

The developer portal describes it as more than 550 active operations across 38 domains. Two HighLevel support articles, on increased scopes for the Anthropic endpoint and on multi-account support for Claude, still state 625 operations across 40 domains, and the first publishes a breakdown of 571 active operations: 243 read, 244 write and 84 delete. Treat every count as a dated snapshot, not a permanent number.

• Scopes decide capability. The operations an assistant sees depend on the OAuth grant, so two connections to the same server can do very different things, and an agent that works in your account can fail in a client account with a narrower grant.
• One sub-account per request. Every request targets a single sub-account. Agency-wide connections across several sub-accounts are still rolling out; if your install offers only a single sub-account, that option is not enabled for your app yet.
• HighLevel’s own pages disagree. The older support-portal setup guide still describes the original /mcp/ endpoint with a Private Integration Token and says OAuth is planned for a future release, and the support articles describe a five-tool design and list OpenAI clients as in development. The developer portal documents OAuth on the per-client endpoints, four tools, and a live OpenAI endpoint. This page follows the developer portal.
• Safety gates. HighLevel says sensitive and irreversible operations require additional confirmation and safety checks.

None of these are reasons to avoid it. They are the operational cost of building on it, and worth pricing in alongside the plan cost covered in our GoHighLevel pricing breakdown.

1. Who operates it? A vendor-run server has a support path and a deprecation policy. A community server on GitHub has neither, however good the code is.
2. How much of the API is reachable? Compare what the agent layer can reach against the vendor’s own published API surface. Do not compare raw tool counts, because a facade design like HighLevel’s deliberately shows a handful of tools in front of a much larger catalog. The number that matters is operations reachable, not tools listed.
3. How does scoping work? Least-privilege scoping is the right answer, but you need to know how it fails and whether the failure is legible to the agent.
4. Is the registry versioned? Tool names are an interface. If they change without notice, agents break in production and the symptom is a confused model rather than a clean error. A versioned endpoint path, like the v2 in HighLevel’s per-client endpoints, is a good sign.