Guide

Connect Advisor OS to Claude, Cursor, and any AI agent with MCP

April 26, 2026 · 8 min read

The Advisor OS MCP server is live. In ten minutes you can plug your pipeline, commissions, contracts, and client data into Claude Desktop, Cursor, ChatGPT, or any agent that speaks the Model Context Protocol. No copy-paste. No screenshots. Just live data, the way your AI tool already wants to read it.

What MCP is, in one paragraph

Model Context Protocol (MCP) is an open standard for connecting AI assistants to the systems where work actually lives. Think of it as USB-C for AI tools: instead of every CRM, ticketing system, and database inventing its own integration, MCP gives them one shared port. Once you connect Advisor OS over MCP, your AI assistant can read your pipeline, search your contracts, query commission data, and look up clients without you ever pasting another spreadsheet into a chat window.

What you can do once it's connected

The Advisor OS MCP server exposes the same data surface your team already uses, scoped to your account. A few of the things that just start working:

"Show me every renewal in the next 90 days that's still in proposal." Returns a clean list, sourced live from your CRM.
"Draft a renewal email for Forge Retail using their actual contract dates." The agent pulls dates and line items from your contracts directly. No paste.
"Summarize this client's spend trend year over year." Ace-style analysis, but inside Claude or Cursor with whatever model you prefer.
"What's my YTD commission by supplier?" Real numbers, real fast.
"Find me a similar deal we won last quarter." Vector search across your historical pipeline.

Prerequisites

An active Advisor OS account on any plan.
An MCP-capable client. Claude Desktop, Cursor, Windsurf, Zed, and the official Anthropic and OpenAI agent SDKs all support MCP today.
An Advisor OS personal access token. Generate one at app.getadvisoros.com/settings/tokens. Treat it like a password.

One token, one scope.

Tokens inherit your role. If you're an advisor, the agent only sees the clients you can already see. If you're an admin, the agent sees what you see. There is no path for the agent to escalate beyond your account permissions.

Connect Claude Desktop

Claude Desktop reads MCP servers from a JSON config file. Open Claude Desktop, go to Settings → Developer → Edit Config, and add Advisor OS to the mcpServers block.

{
  "mcpServers": {
    "advisor-os": {
      "command": "npx",
      "args": ["-y", "@advisor-os/mcp"],
      "env": {
        "ADVISOR_OS_TOKEN": "aos_live_xxxxxxxxxxxxxxxxxxxx"
      }
    }
  }
}

Save the file and restart Claude Desktop. You'll see a hammer icon in the input box. Click it; "advisor-os" should appear with green status. Type "list my open deals" and Claude will route the request through the server and return live data.

Connect Cursor

Cursor uses the same JSON shape under Cursor Settings → MCP → Add new MCP server. You can also drop a .cursor/mcp.json into any project to scope the connection per-repo.

{
  "mcpServers": {
    "advisor-os": {
      "command": "npx",
      "args": ["-y", "@advisor-os/mcp"],
      "env": {
        "ADVISOR_OS_TOKEN": "aos_live_xxxxxxxxxxxxxxxxxxxx"
      }
    }
  }
}

Restart Cursor. The server shows up in the agent panel. Cursor will ask once, the first time the model wants to call a tool, whether to allow it. Approve, and from then on Cursor remembers the choice for that workspace.

Connect any agent over HTTP

If you're building your own agent, integrating a workflow tool like n8n, or wiring up the OpenAI or Anthropic Agent SDKs, point at the hosted HTTP endpoint instead of running the server locally:

POST https://mcp.getadvisoros.com
Authorization: Bearer aos_live_xxxxxxxxxxxxxxxxxxxx
Content-Type: application/json

The endpoint speaks both JSON-RPC 2.0 and the streaming SSE transport. Tool discovery, schema validation, and pagination are all handled by the server. The full tool list, schemas, and example payloads are in the MCP developer docs.

The tools the server exposes

Each tool maps to a concrete read or write operation in Advisor OS. The agent picks the right one based on the user's question. You don't have to memorize anything; the names are self-describing.

list_deals, get_deal, create_deal, update_deal
list_clients, get_client, get_client_spend
list_contracts, get_contract, find_renewals
list_commissions, commission_summary
search (universal search across deals, clients, contracts, partners)
list_partners, get_partner
create_note, create_task

Three workflows worth setting up first

1. The Monday morning pipeline review

Open Claude or Cursor, type "give me a pipeline summary by stage with deals over $10k MRC". The model calls list_deals, groups, and returns a clean table. Paste it into your standup notes and move on.

2. The "what is this client paying for again" lookup

Mid-call, no time to dig. Ask "what active contracts does Acme Corp have, and what's their MRC?" The model calls list_contracts filtered by client and gives you the answer in three seconds. Faster than opening the CRM.

3. Renewal outreach drafts

Ask "draft a renewal email for any client whose contract ends in the next 60 days, using their service mix and our standard renewal language." The model finds the renewals, reads each contract, and produces a draft per client. Approve, edit, send.

Security and what the server cannot do

We took a deliberately conservative posture for this launch.

Read-by-default, write-by-approval. Tools that change data (create deal, update contract, send notification) are off by default. You enable them per-token in settings.
Scoped to your role. The token cannot see beyond your existing role permissions. Advisor tokens can't read other advisors' books. Client-admin tokens can't see internal commission data.
No training on your data. Advisor OS does not send your data to third-party model providers for training. The server is a transport; your AI client is responsible for its own model policies.
Audit log. Every tool call is logged with the token ID, tool name, and timestamp. You can see them in the audit log at any time.
Revoke instantly. Delete the token from settings and the connection dies on the next request.

Troubleshooting

The server doesn't appear in Claude Desktop. Restart the app; Claude only re-reads the config on launch. Check the file path is right for your OS (macOS: ~/Library/Application Support/Claude/claude_desktop_config.json, Windows: %APPDATA%\Claude\claude_desktop_config.json).

"Authentication failed" on every tool call. Regenerate the token. Tokens are scoped to a single user; if the user's role changed, the old token may have been invalidated.

Tool calls hang. The agent's network layer may be blocking outbound requests. Check that your firewall allows mcp.getadvisoros.com on 443.

Empty results when you expect data. Ask the model to explain its tool call ("what filter did you pass to list_deals?"). The most common cause is an over-narrow date range or a stage filter that excludes everything.

Where to go from here

The MCP docs at docs.getadvisoros.com/mcp have the full tool reference, response schemas, and examples for the most common workflows. If you're building something interesting on top of the server, we'd love to hear about it. Get in touch.