AI & Automation May 23, 2026 8 min read By Agni Folio Team

Connect Your Own AI to Agni Folio: The MCP Server Is Live

You already pay for Claude. You already use Cursor. You already have a local model running on your workstation. Why should your portfolio data live in one more chat window, walled off from the AI you actually work with every day?

From today, it doesn't have to. Agni Folio has shipped a Model Context Protocol (MCP) server at https://agnifolio.com/mcp. Point any MCP-compatible AI client at that URL, sign in once, and your AI can read your holdings, run analytics on your portfolio, and (if you grant the scope) propose buys, sells, and dividend records — directly from the chat window you already use.

This post explains what we shipped, why we shipped it, how to connect in two minutes, and how we kept it safe.

What "Connect Your AI" actually means

For the last year, the most useful AI feature inside Agni Folio has been Astra — a chat agent with read access to your portfolio and a confirm-before-mutate path for adding or editing entries. Astra is great. But it lives inside our web app, and you have to switch tools to use it. If you've already got Claude Desktop open, or you live in Cursor all day, that context switch is friction.

The Model Context Protocol is an open spec — backed by Anthropic, adopted by Cursor, Cline, Continue, and a growing list of AI tools — that lets any compliant client call typed functions on remote servers. It's the same idea as a REST API, but designed for LLMs: tools are discovered dynamically, described in plain English, and called over JSON-RPC.

Our MCP server exposes the same tool registry Astra uses internally. That means:

  • Anything Astra can read inside the app, your external AI can also read.
  • Anything Astra can change (with confirmation) inside the app, your external AI can also propose — under a scoped, revocable token.
  • Every call is audit-logged in the same place as Astra's, and shows up in your Connected AI settings.

One backend, two surfaces. Less code, fewer surprises.

Why this matters more than it sounds

"Add an MCP server" reads like a checkbox feature. It isn't. Three things change once your AI can see your real portfolio without copy-paste:

1. Your AI stops guessing about you

Every conversation you've ever had with a general-purpose chatbot about your money started with the same overhead: pasting in a sanitised version of your holdings, hoping you didn't typo a ticker, hoping the model didn't confabulate your cost basis. With MCP, the model fetches the data itself, with the actual current numbers. The quality of the conversation goes up immediately.

2. Cross-tool workflows finally work

The most powerful pattern in AI right now isn't one chat window — it's chaining context across many. Your AI can pull from Agni Folio, your calendar, your code repo, and your research notes in a single conversation. "Look at my Agni Folio holdings, compare them to the tickers in this Google Doc, and tell me which I'm under-allocated on" used to be a Sunday-afternoon spreadsheet job. Now it's a sentence.

3. You can use the model you want, on the hardware you want

Locked into one vendor's chat UI is no longer the cost of getting AI-assisted portfolio analysis. Run Claude, GPT-4, a local Llama, a fine-tuned domain model — whatever you trust most — and connect it to Agni Folio with the same standard interface. If a better model comes out tomorrow, you swap the client; the integration stays the same.

How to connect in two minutes

The fast path:

  1. Open your MCP-enabled AI client (Claude Desktop, Cursor, Cline, Continue, or any other).
  2. Add a remote MCP server with the URL https://agnifolio.com/mcp.
  3. The client will discover our OAuth flow automatically. You'll be redirected to an Agni Folio consent screen.
  4. Approve the scopes you want to grant — agnifolio:read for read-only access, agnifolio:write for the ability to propose changes.
  5. Done. Ask your AI: "List my Agni Folio holdings and tell me my technology-sector concentration."

The full step-by-step, with client-specific config snippets for Claude Desktop and Cursor, lives in the Connect Your AI (MCP) docs page.

What you can ask your AI to do

Anything Astra can do. Examples we've tested:

Analytics

  • "What's my total net worth across all currencies, in USD?"
  • "Show me my dividend income for the last 12 months, broken down by holding."
  • "Which of my positions has the largest unrealised loss right now?"
  • "What's my technology-sector exposure and how does it compare to global market cap weighting?"

These are agnifolio:read-scope calls. The model picks the right tool, fetches data, formats an answer.

Scenario modelling

  • "If I sell half my largest position at today's price, how does my asset allocation change?"
  • "Hypothetically, if I added $10,000 of crypto, what would my new allocation look like?"

These produce previews without modifying anything.

Mutations (with confirmation)

  • "Add a buy transaction: 50 shares of MSFT at $410 in my brokerage account, dated yesterday."
  • "Record a dividend of $124.50 received today on my AAPL position."
  • "Delete the duplicate ICICIBANK entry in my Indian brokerage account."

These require agnifolio:write scope. The model proposes the action, you (or the AI client's confirmation UI) approve, then it executes.

Bulk import inside your AI

Paste a broker statement into your AI chat and say "add these to my Agni Folio." Your AI parses the statement, maps tickers to your existing entries, proposes the transactions, and waits for your sign-off. The same import flow that's been the most-loved feature inside Astra — now from any AI client.

The controls you actually have

Connecting external AI to a portfolio is the kind of thing you should be a little suspicious about. Here is what we built so you stay in control.

OAuth 2.1, no shared secrets

The handshake follows the OAuth 2.1 authorisation-code flow. Your AI client never sees your password, never sees your Google sign-in, never holds a long-lived credential it could lose. It holds a scoped, revocable access token, and only that.

Dynamic Client Registration

We don't maintain a hard-coded list of "allowed" AI clients. Per RFC 7591, any client self-registers and you approve it by name on the consent screen. If a new MCP client launches tomorrow, it works on day one — no engineering update required from us.

Read or write — your choice

You decide at the consent screen whether to grant agnifolio:read only, or also agnifolio:write. Most people start read-only, see how their AI behaves for a few sessions, and only upgrade if they want it to actually add transactions on their behalf.

Audit trail per call

Every external MCP call is logged with the client name, the tool that ran, and when. You can see the trail of your own AI's activity from Settings → Connected AI. If you ever wonder "did my AI actually do that?" — open the log.

Revoke any client in one click

The same settings page lists every AI client you've connected. One click revokes access. The next call from that client fails immediately. No cooldown, no support ticket, no waiting.

What about pricing?

The MCP server is part of Agni Folio's free-forever core. There is no usage cap, no Pro tier, no add-on for this feature. If you can use Astra inside the app today, you can connect your own AI today. We may someday introduce optional paid features built on top of MCP (think: long-running automated workflows, scheduled rebalancing alerts), but everything that works in the app today via Astra will also work via your own AI, free.

Who this is for

  • Heavy AI users. If you already pay for a Claude Pro or Cursor subscription, MCP lets you use that investment for portfolio analysis instead of paying for two separate AI features.
  • Privacy-leaning investors. Connect a local LLM via Cline or Continue and your portfolio analysis happens on-device after the data is fetched. Useful if you'd rather not send holdings to a hosted model.
  • Power users. Custom MCP clients can drive scripts, dashboards, and automations. Anything that speaks the spec works.
  • Cross-tool builders. Anyone running multi-source AI workflows — research + portfolio + tax + calendar — gets a first-class portfolio source in their toolkit.

Who this is not (yet) for

  • Investors who want real-time market-hour quotes — our MCP returns cached daily prices, same as the app.
  • Anyone looking for tax-filing-grade reporting — Astra and MCP can summarise gains, but you still need a CPA for filings.
  • Users without an MCP-compatible client — most mainstream chat apps don't support MCP yet. Astra inside the app remains the best entry point for them.

Start using it

The MCP server is live now. To connect:

If your AI client has a list of "MCP servers worth connecting to," we hope to earn a spot. If you build something interesting on top of this — a dashboard, a script, a multi-source workflow — we'd love to see it. Let us know.