Connect Your AI (MCP)

Agni Folio ships a Model Context Protocol (MCP) server. Point any MCP-compatible client — Claude Desktop, Cursor, Cline, Continue, or your own — at our server, sign in once, and your AI can read your holdings, run portfolio analytics, and propose transactions just like Astra does inside the app.

Tip

The MCP server reuses the same tool registry Astra uses inside Agni Folio. Anything Astra can do, your AI can do — with the same per-user scoping, the same audit trail, and the same confirm-before-mutate safety.

What it is

Model Context Protocol is an open spec for connecting AI assistants to external data and tools. Instead of pasting your portfolio into ChatGPT or building bespoke integrations for every tool, MCP gives any AI client a standard way to call typed functions on your data.

Agni Folio’s MCP server lives at:

https://agnifolio.com/mcp

It speaks JSON-RPC 2.0, supports MCP protocol version 2025-03-26, and is gated by OAuth 2.1 with PKCE.

Why you might want this

• Use the AI you already pay for. If you have Claude Desktop, Cursor, or a Pro plan elsewhere, you don’t need to use Astra inside the Agni Folio web app — your existing AI can answer portfolio questions directly with your real data.
• Bring your own model. Run a local LLM via an MCP-compatible client (Cline, Continue) and keep your portfolio analysis fully on-device after the data is fetched.
• Mix it into wider workflows. Ask your AI to “look at my Agni Folio holdings and compare them to the watchlist in this Notion doc” — MCP lets one assistant pull from many sources in a single conversation.
• Build automations. Any tool that speaks MCP — including custom scripts — can query your portfolio without screen-scraping or unofficial APIs.

Supported clients

Any client that implements MCP 2025-03-26 with OAuth 2.1 should work out of the box. Tested:

Client	Status
Claude Desktop	✅ Supported via custom MCP server URL
Cursor	✅ Supported via mcpServers config
Cline	✅ Supported (VS Code extension)
Continue	✅ Supported (open-source AI coding assistant)
Custom MCP clients	✅ Anything implementing the MCP 2025-03-26 spec

How to connect

1. Open the MCP settings in your AI client.

   Each client has its own way to add a remote MCP server. Look for a “Remote MCP server” or “Custom MCP server URL” field.

   Claude Desktop

   In Settings → Developer → MCP Servers, click “Add” and enter:

   https://agnifolio.com/mcp

   Claude Desktop will automatically discover the OAuth flow via .well-known/oauth-authorization-server and walk you through the rest.

   Cursor

   Open ~/.cursor/mcp.json (or Settings → MCP) and add:

   {
     "mcpServers": {
       "agnifolio": {
         "url": "https://agnifolio.com/mcp"
       }
     }
   }

   Restart Cursor; it will prompt you to authorize on first use.

   Cline / Continue / others

   Use whatever “add remote MCP server” UI your client provides, with URL:

   https://agnifolio.com/mcp

   Any client implementing the MCP 2025-03-26 spec and OAuth 2.1 dynamic client registration will work.

2. Approve the OAuth consent screen.

   The first time your AI client calls Agni Folio, you’ll be redirected to a consent screen showing:

   • Which client is requesting access (by name)
   • Which scopes it wants: agnifolio:read, agnifolio:write, or both
   • The plain-English permissions those scopes grant

   Click Approve to grant access, or Deny to abort. We never share your credentials with the AI client — only a scoped access token.

3. Start asking your AI portfolio questions.

   Open your MCP-enabled chat and try:

   • “List my Agni Folio holdings”
   • “What’s my exposure to technology stocks?”
   • “Add a buy transaction for 100 shares of AAPL at $185 to my brokerage account” (mutating — requires agnifolio:write scope)

   Your AI will discover the available tools, call the right one, and answer with your real data.

What your AI can do

Your connected AI gets the same tool catalogue Astra has inside Agni Folio:

Capability	Tool examples	Scope required
Read holdings	list_holdings, get_entry_details, get_recent_transactions	agnifolio:read
Run analytics	compute_concentration, compute_xirr, get_dividend_summary	agnifolio:read
Propose changes	add_portfolio_entry, add_transaction, edit_entry, delete_entry	agnifolio:write

Use tools/list (the standard MCP discovery method) to see the live, up-to-date catalogue from any client.

Scopes — read vs write

Scope	Grants	Recommended for
agnifolio:read	Read holdings, run analytics, view transactions	Read-only AI assistants, dashboards, research
agnifolio:write	Everything in read PLUS create / edit / delete entries and transactions	AI assistants you trust to make changes on your behalf

You decide which scopes to grant at the consent screen. Most users start with read only and upgrade to write once they’ve seen how the AI behaves.

Security model

• OAuth 2.1 + PKCE. Industry-standard authorization-code flow — no password sharing, no copy-paste tokens.
• Dynamic Client Registration (RFC 7591). New MCP clients self-register; you approve them by name on the consent screen.
• Scope-gated calls. Read-only tools require agnifolio:read; mutating tools require agnifolio:write. You choose which to grant on the consent screen.
• Per-call audit trail. Every tool execution is logged: which client called, which tool, when. You can audit your own usage from Settings → Connected AI.
• Revocable at any time. Disconnect a client from Settings → Connected AI. Future requests from that client are rejected immediately.

Audit trail and managing connected AI

Open Settings → Connected AI to:

• See every AI client connected to your account, by name
• Review the scopes each one has
• See the last time each client called the server
• Revoke any client with one click (instant — no cooldown)

Every external MCP call shows up alongside Astra’s in-app activity in the same audit view.

OAuth discovery (for client developers)

If you’re building or debugging an MCP client, our discovery endpoints follow the spec:

# Authorization server metadata (RFC 8414)
curl https://agnifolio.com/.well-known/oauth-authorization-server

# Protected resource metadata
curl https://agnifolio.com/.well-known/oauth-protected-resource

Both return JSON declaring our authorization_endpoint, token_endpoint, registration_endpoint, revocation_endpoint, supported scopes, supported response types, and code_challenge_methods_supported: ["S256"].

The resource URL is https://agnifolio.com/mcp.

When MCP is the wrong tool

• Real-time price quotes during market hours. Agni Folio’s MCP returns cached daily prices, not the live tape. Use your broker’s app for execution-time prices.
• Bulk one-time imports. If you have a 500-row broker statement, the dedicated AI Transaction Import page is faster and more deterministic than asking an external AI to do the same.
• Sensitive accountability conversations. Money Crew check-ins, legacy nominee setup, and other people-facing workflows live inside the Agni Folio app — they’re not exposed via MCP.

What’s next

• Astra AI Agent — the same toolset, surfaced inside the Agni Folio web app instead of via your external AI
• Portfolio Insights — the static, in-app analytics view
• Tracking Investments — how the underlying data your AI will be reading is structured