---

Last updated October 26, 2026

Wealthfolio’s MCP server lets external AI agents talk to your portfolio over the Model Context Protocol. Point Claude Code (including Claude Desktop’s Code tab), Codex, the ChatGPT desktop app, Cursor, Windsurf, Cline, or any MCP-capable client at Wealthfolio and it can read your holdings, performance, and activities — and, if you allow it, draft or record new activities — through the exact same typed tools the built-in AI Assistant uses.

Every connection is authenticated with a scoped access token you create and can revoke. Nothing is exposed until you turn the server on and hand out a token.

---

1 · Where it lives

Open Settings → AI Agent Access. What you see depends on how you run Wealthfolio:

• Desktop app — a master toggle turns the feature on, then a local MCP server runs at http://127.0.0.1:<port>/mcp. You control it with Start/Stop, choose whether it starts with the app, and toggle audit logging.
• Self-hosted / web server — the endpoint is enabled with an environment variable (see §6) and served at /mcp on the same origin as your server. Token management and the audit log work the same way.

---

2 · Turn it on and create a token

1. Enable the server. Flip the master toggle (desktop) or set the environment variable (self-hosted). On desktop, press Start — the endpoint URL appears with a green dot when it’s live.
2. Create an access token. Under Personal access tokens, click Create token, name it (e.g. “Claude Code”), pick an expiry, and choose its scopes. Tokens default to read-only — one click away from the common case.
3. Copy the token once. It’s shown a single time at creation. The dialog also gives you ready-to-paste setup for Claude Code, Codex, or JSON-config clients, with the token already embedded.

Expiry options are 7 days, 30 days, 90 days (default), 1 year, or no expiry. Shorter is safer — you can always mint a new one.

---

3 · Scopes — what a token can do

A token carries a set of scopes. The agent can only call tools its scopes allow; the backend rejects anything outside them. Scopes split into read and write.

Read access

Write access

The dialog offers four presets that cover most needs: Read-only, Read + draft, Read + write, and Read + write + suggest (which grants both classification suggest and write). Activities — commit/write always pulls in Activities — draft, and Classification — write pulls in Classification — suggest — an agent has to be able to stage a change before it can commit one.

---

4 · Connect a client

Wealthfolio speaks Streamable HTTP MCP. The server card’s Copy config button copies the Claude Code command below with YOUR_TOKEN as a placeholder. Its help (?) button and the token dialog have a Claude Code, Codex / ChatGPT and Other tab with the snippet for each app; the token dialog fills in your new token. Replace PORT with the port shown on the server card (self-hosted users use their server’s URL) and YOUR_TOKEN with the token you copied.

Wealthfolio labels each tool as read-only or as a write, and marks the writes that change or remove existing data. Clients use these labels to decide when to ask you first: Codex, for example, runs read tools without asking and asks before tools that change existing data. Draft and suggest tools count as read-only: they return a preview and save nothing.

Claude Code and Claude Desktop (Code tab)

Run this in a terminal:

claude mcp add --scope user --transport http wealthfolio http://127.0.0.1:PORT/mcp --header "Authorization: Bearer YOUR_TOKEN"

--scope user adds Wealthfolio for all your projects (Claude Code keeps it in ~/.claude.json). The Code tab in Claude Desktop runs Claude Code, so it picks up the same server. Run claude mcp list, or /mcp in a session, to check the connection.

To share the server with one project instead, add a .mcp.json at the project root. It reads the token from the WEALTHFOLIO_TOKEN environment variable, so the file holds no secret and is safe to commit:

{
  "mcpServers": {
    "wealthfolio": {
      "type": "http",
      "url": "http://127.0.0.1:PORT/mcp",
      "headers": { "Authorization": "Bearer ${WEALTHFOLIO_TOKEN}" }
    }
  }
}

Set WEALTHFOLIO_TOKEN in the shell that starts Claude Code. Claude Code asks you to approve a project’s servers the first time it uses them.

Codex and the ChatGPT desktop app

The Codex CLI, the Codex IDE extension, and the ChatGPT desktop app share one config file, ~/.codex/config.toml. Add:

[mcp_servers.wealthfolio]
url = "http://127.0.0.1:PORT/mcp"
http_headers = { Authorization = "Bearer YOUR_TOKEN" }
tool_timeout_sec = 120

To keep the token out of the file, remove Authorization from http_headers and add bearer_token_env_var = "WEALTHFOLIO_TOKEN" instead. Codex then reads the token from that environment variable. Run codex mcp list to check the server is configured.

Other apps

The app’s Other tab has a snippet for each of these. Cursor, Windsurf, Cline and Unsloth Studio share one config shape:

{
  "mcpServers": {
    "wealthfolio": {
      "type": "http",
      "url": "http://127.0.0.1:PORT/mcp",
      "headers": { "Authorization": "Bearer YOUR_TOKEN" }
    }
  }
}

LM Studio uses the same shape without the "type" line.

VS Code differs — it uses a top-level servers key:

{
  "servers": {
    "wealthfolio": {
      "type": "http",
      "url": "http://127.0.0.1:PORT/mcp",
      "headers": { "Authorization": "Bearer YOUR_TOKEN" }
    }
  }
}

OpenCode uses its own mcp key. "oauth": false stops it from starting an OAuth flow for this server:

{
  "mcp": {
    "wealthfolio": {
      "type": "remote",
      "url": "http://127.0.0.1:PORT/mcp",
      "oauth": false,
      "headers": { "Authorization": "Bearer YOUR_TOKEN" }
    }
  }
}

Jan has its own format too; pick Jan in the app’s Other tab to copy it. Any other client that speaks Streamable HTTP just needs the URL plus an Authorization: Bearer YOUR_TOKEN header.

Where the config goes

On a self-hosted server with several profiles, the snippets the app gives you also carry an X-WF-Profile-Id header for the profile you’re signed in to. Keep it; without it the server uses the default profile.

Once connected, the agent discovers Wealthfolio’s tools automatically. Ask it “what are my top holdings?” or “how did my portfolio do this year?” and it’ll call the read tools and answer with your real numbers.

Use your Claude or ChatGPT subscription

Claude Code (in a terminal, an IDE, or Claude Desktop’s Code tab) signs in with your Claude plan, and Codex (CLI, IDE extension, or the ChatGPT desktop app) signs in with your ChatGPT plan. Connected to Wealthfolio, they use that plan: you don’t need an API key.

Some Claude apps can’t reach a local Wealthfolio:

• claude.ai and the Claude mobile apps connect to MCP servers from Anthropic’s cloud, which can’t see 127.0.0.1 on your computer.
• The Claude Desktop chat tab only starts local command servers, not HTTP ones (see above).

The community mcp-remote package can wrap an HTTP server as a local command for the Claude Desktop chat tab. It needs Node.js, and Wealthfolio doesn’t test or support it.

The in-app AI Assistant is different: it calls a model provider directly, so it needs an API key or a local model.

---

5 · Audit log

Every tool call an agent makes is recorded under Agent activity — the time, the tool name, the outcome (Success, Denied, or Error), and which token made the call. Filter by tool or outcome, search, and Clear log when you want a clean slate.

On desktop, the Log agent activity switch turns recording on or off. On a self-hosted server it’s controlled by an environment variable (below). With logging off, no new audit rows are written.

---

6 · Self-hosted servers

On a self-hosted Wealthfolio server the MCP endpoint is off by default and enabled with environment variables:

After enabling, create tokens from Settings → AI Agent Access exactly as on desktop. The url in your client config is your server’s origin plus /mcp (for example https://wealth.example.com/mcp).

---

7 · Security model

• Nothing is exposed by default. The server is off until you enable it, and it serves nothing without a valid token.
• Local-only on desktop. The desktop server binds to 127.0.0.1 — it isn’t reachable from other machines on your network.
• Scoped and revocable. Each token carries only the scopes you granted. Remove a token and any client using it loses access immediately.
• Writes are gated. Committing an activity requires both the draft and write scopes; a draft-only token can stage changes but never commit them.
• Stopping ≠ revoking. Stopping the server (or disabling the feature on desktop) keeps your tokens valid — they resume working when you start it again. To cut a client off for good, delete its token.
• Typed tools, not raw SQL. Agents call the same scoped, typed tools as the AI Assistant. They never receive your database or your other API keys.

---

8 · Troubleshooting

The client can’t connect

• Make sure the server is running (green dot on the server card, or WF_MCP_ENABLED=true and a restart on self-hosted).
• Check the url and PORT match what’s on the server card, and that the path ends in /mcp.
• Confirm the Authorization header is Bearer <token> with a valid, unexpired token.

The container won’t start after setting WF_MCP_ENABLED=true

On self-hosted installs that bind to the network, MCP requires Wealthfolio auth. If you set WF_AUTH_REQUIRED=false and rely only on a reverse proxy for login, the server refuses to start because the Agent Access API can mint long-lived MCP tokens.

Fix it by configuring WF_AUTH_PASSWORD_HASH or OIDC, then restart with WF_MCP_ENABLED=true. If you don’t want Wealthfolio auth enabled, leave WF_MCP_ENABLED=false.

The agent says a tool was “denied”

The token is missing the scope for that tool. Create a new token with the scope you need (or the right preset) — scopes can’t be edited after creation, so mint a fresh one and remove the old.

It worked yesterday, not today

Check whether the token expired (see its status in the token list) or was removed. On desktop, also confirm the server is started and, if you rely on it, that auto-start is on.

---

Frequently asked questions

Does Wealthfolio have a built-in MCP server?

Yes. Wealthfolio ships a native Model Context Protocol server — there’s no third-party bridge to install. Turn it on under Settings → AI Agent Access, mint a scoped token, and point any MCP client at it. On the desktop app it runs locally at http://127.0.0.1:<port>/mcp; on a self-hosted server it’s served at /mcp on your own origin.

Which MCP clients work with Wealthfolio?

Any MCP client that speaks Streamable HTTP, including Claude Code (and Claude Desktop’s Code tab), Codex, the ChatGPT desktop app, Cursor, Windsurf, Cline, and the VS Code MCP integration. The Claude Desktop chat tab, claude.ai, and the Claude mobile apps can’t reach a local Wealthfolio; see Use your Claude or ChatGPT subscription.

Is the Wealthfolio MCP server safe?

Access is off until you enable it, and every connection uses a scoped, revocable access token you create yourself. Agents reach your data only through the same typed tool registry the in-app AI Assistant uses — never your raw database — and every call is recorded in an audit log you can review.

---

Next step: If you’d rather chat with your portfolio from inside the app instead of wiring up an external agent, see the AI Assistant.

---