Resources · AI workbench

Cursor, Claude Desktop and MCP

How I set up desktop AI to build and operate HighLevel systems — and how that connects to the hub orchestration model we use on every agent build.

Two tools, two jobs

What I actually use them for

Both speak MCP. Neither replaces Super Agents inside HighLevel — they sit in the operator layer, above the hub.

Cursor

IDE with agent mode — marketplace apps, MCP server code, integrations, this website. Codebase-aware, terminal access, project rules.

Use for: anything that ships as files in git.

Setup guide (MD)

Claude Desktop

Desktop app with MCP connectors — architecture drafts, KB review, scoping from transcripts, integration design when you are not in the repo.

Use for: thinking, documents and MCP-connected research.

Setup guide (MD)

Methodology

Hub orchestration — what that actually means

Orchestration is not picking a chatbot. It is how rules, judgment and operators move data through HighLevel as the communications hub.

Three-layer orchestration: Cursor and Claude Desktop as operator layer, Super Agents as judgment, workflows as rules, HighLevel hub at centre with MCP bridge

Download SVG

Layer A — Rules

Workflows, n8n, webhooks. Fixed logic. No improvisation.

Layer B — Judgment

Super Agents, Voice AI. Skills + KB + MCP tools + handoff.

Layer C — Operator

Cursor, Claude Desktop. Build and maintain the stack — not customer-facing.

Every build — including AI — follows the same four phases as client work: audit, stage, apply, hand back. You keep the account and the data; nothing touches live traffic until you sign off the phase.

Download full methodology (MD)

MCP setup

Connect desktop AI to your systems

Model Context Protocol gives Claude and Cursor typed, permissioned tools — not a scraped guess at your CRM. Start read-only on a staging sub-account.

MCP config starter (JSON)

Example claude_desktop_config.json structure — filesystem + placeholder HighLevel staging server.

Download

Claude Desktop MCP setup

Config file paths for Windows and Mac, restart steps, troubleshooting.

Download

Cursor MCP setup

Project rules, .cursor/mcp.json, servers worth connecting for HighLevel builds.

Download

What MCP is for: HighLevel contact lookup, calendar slots, custom middleware, Fathom transcripts, Supabase rows — each as a separate bounded server. I build these for production agents; desktop MCP uses the same pattern in staging.

Production agents

Desktop AI is not the same as Super Agents

Use Cursor and Claude Desktop to design agents. Deploy customer-facing intelligence inside HighLevel with the skill library, prompt templates and production MCP servers scoped to one job.

Need it built end-to-end? That is Super Agents plus MCP & APIs — staged the same way as everything else.