MCP Server

Treeline includes a built-in MCP (Model Context Protocol) server. It lets AI agents query your finances, sync accounts, tag transactions, and more.

It speaks MCP two ways: over STDIO for an agent on the same machine (tl mcp), and over HTTP for a networked agent that connects to a hub you run (tl serve). The tools are the same either way.

Setup

On your machine (STDIO)

For an agent running on the same machine as your data — it launches tl mcp itself. Pick the client you’re using; you can use more than one, and they all share the same local database.

Claude Desktop

Two ways to connect, depending on whether you use the terminal.

Don’t use the terminal? Download the .mcpb extension from treeline.money/download and open it. Claude Desktop installs it automatically — no configuration needed.

Already have the CLI? Open Settings > Developer > Edit Config in Claude Desktop and add:

{
  "mcpServers": {
    "treeline": {
      "command": "tl",
      "args": ["mcp"]
    }
  }
}

Restart Claude Desktop. You should see Treeline listed in the tools menu.

Which should I choose?

The extension bundles its own copy of the CLI, so to get new features you download a new extension. The CLI config uses tl from your PATH — run tl update and Claude Desktop picks up the new version on next restart. If you already have the CLI installed, the CLI config is the better choice.

Both options work in Chat and Cowork mode.

Claude Code

Requires the Treeline CLI.

claude mcp add treeline -- tl mcp

Gemini CLI, Cursor, and others

Requires the Treeline CLI. Any client that supports the MCP STDIO transport can connect — point it at tl mcp as the command.

Over the network (a hub)

A hub is any Treeline instance running tl serve, and it does two things:

• Replication — your other devices push and pull the database to and from it, so your data can live on more than one machine.
• Remote MCP — it exposes this MCP server over HTTP so a networked agent can reach it.

(See What’s a Hub? for the full picture.)

This page is about the second role. An agent that can’t launch a local process — ChatGPT, Claude on your phone, or any client that supports a remote (custom) MCP server — connects to your hub’s /mcp endpoint and makes requests over the network. Unlike push/pull, your data isn’t copied to the client — it stays on the hub, and the agent receives only the results of the queries it runs. The hub walks the client through authorizing with the scopes you approve.

Set up a hub →

Available Tools

Once connected, the agent has access to these tools:

Tool	Description	Access
status	Account balances, transaction counts, and connected integrations	Read-only
schema	Get database schema (tables, views, columns) for SQL query planning	Read-only
query	Execute a read-only SQL query against your financial database	Read-only
query_write	Execute a SQL query with write access (INSERT, UPDATE, DELETE)	Read/Write
sync	Pull latest transactions from connected bank integrations	Read/Write
tag	Apply tags to transactions	Read/Write
doctor	Run database health checks	Read-only
demo	Enable, disable, or check demo mode	Read/Write
skills_list	List available user-created financial skills	Read-only
skills_read	Read a file from a skill directory	Read-only
skills_write	Create or update a skill file	Read/Write
encryption_status	Check database encryption and lock status	Read-only
version	Get current CLI version and check for updates	Read-only

Over a hub the same tools are available, but the ones that read your database return an error until you’ve pushed one to it (tl hub push).

Usage

You don’t need to mention Treeline or MCP by name. Just ask questions naturally:

• “What’s my net worth?”
• “How much did I spend on groceries last month?”
• “Show me my largest transactions this week”
• “Tag all Starbucks transactions as coffee”
• “Sync my bank accounts”

The agent will choose the appropriate Treeline tool based on your question.

Privacy

Where your data goes depends on how the agent connects:

• STDIO (same machine): the server runs on your machine and reads your Treeline database on disk. Nothing is sent anywhere by the server itself.
• Over a hub (HTTP): the server runs on your hub. A networked agent reaches it over HTTPS and gets back only the results of the queries it runs — your full database stays on the hub, not on the client.

Either way, the agent — and the LLM behind it — sees the query results needed to answer you. With a cloud LLM, those results reach the provider (the same as any cloud assistant); with a local model, nothing leaves. The provider sees what you ask about, not your whole database.