actual-mcp

actual-mcp

Enables agents to read and edit budgets in Actual Budget via natural language, with a safety layer that requires confirmation for destructive actions.

Category
Visit Server

README

actual-mcp

CI License: MIT

An MCP server for Actual Budget — with a per-tool safety layer.

Actual has no HTTP or MCP surface of its own; it ships a Node SDK. This wraps that SDK as an MCP server so any agent — Claude Desktop, Cursor, your own — can read and edit your budget in natural language.

The difference from a plain wrapper: every tool carries its own risk. A read is a read. A mutation says so. And a delete can never fire on the agent's say-so — it returns a confirmation token and refuses until you call it again with confirm: true. The model can plan freely; it cannot quietly wipe a category because a prompt drifted.

Giving an agent write access to your finances is a liability unless the tools themselves carry the seatbelt. The safest place to enforce a rule is the tool, not the prompt.


Requirements

  • Node.js 20+
  • A running, self-hosted Actual sync server (the actual-server you sync your budget to — not the desktop app). You need its URL, its password (if set), and the Sync ID of the budget you want to expose.

Quick start

1. Find your Sync ID

In Actual: Settings → Show advanced settings → Sync ID. It's a UUID like a1b2c3d4-....

2. Add the server to your MCP client

<details open> <summary><b>Claude Desktop</b></summary>

Edit claude_desktop_config.json (Claude → Settings → Developer → Edit Config):

{
  "mcpServers": {
    "actual": {
      "command": "npx",
      "args": ["-y", "actual-mcp"],
      "env": {
        "ACTUAL_SERVER_URL": "https://actual.example.com",
        "ACTUAL_PASSWORD": "your-server-password",
        "ACTUAL_SYNC_ID": "your-sync-id"
      }
    }
  }
}

Restart Claude Desktop. You'll see the tools appear under the 🔌 icon. </details>

<details> <summary><b>Cursor</b> / other MCP clients</summary>

Any stdio MCP client works — point it at npx -y actual-mcp with the same four env vars. In Cursor: Settings → MCP → Add new and use the same command/args/env shape. </details>

3. Ask it things

What did I spend on groceries last month?
List my accounts and their balances.
Move R$200 from Dining to Savings for July.
Add a R$54.90 transaction to Nubank, payee "Uber", category Transport, dated today.

Configuration

Env var Required What
ACTUAL_SERVER_URL Your Actual sync server URL.
ACTUAL_SYNC_ID Sync ID of the budget to open.
ACTUAL_PASSWORD Server password. Omit if the server has none.
ACTUAL_DATA_DIR Where the SDK caches the budget. Defaults to /tmp/actual-mcp-data.

Amounts are integer minor units, matching the Actual SDK — R$200.00 is 20000, and outflows are negative (-5490 = R$54.90 spent).

The safety model

Level Tools Behaviour
🟢 read-only all list_*, get_* Never touch data.
🟡 mutating all create_*, update_*, set_budget_* Change data; run immediately.
🔴 guarded all delete_* Change data; refuse until confirm: true.

Every tool's level is exposed three ways so the seatbelt is visible before a call: in the tool description, in MCP readOnlyHint / destructiveHint annotations, and — for guarded tools — as a hard refusal at call time.

What a guarded call looks like. The agent tries to delete, and gets back a refusal with the exact arguments and a confirmation token:

// → delete_category { "id": "cat-abc-123" }
// ← (isError)
⚠️ "delete_category" is a destructive action and needs confirmation.
It will run with:
{ "id": "cat-abc-123" }
To proceed, call "delete_category" again with "confirm": true.
Confirmation token: 75aa4f772b62

Only an explicit re-invocation clears it:

// → delete_category { "id": "cat-abc-123", "confirm": true }
// ← { "ok": true }

The human reading the transcript sees precisely what "yes" authorized.

Tools (28)

Tool Level Args
list_category_groups 🟢
create_category_group 🟡 name
update_category_group 🟡 id, name
delete_category_group 🔴 id, confirm
list_categories 🟢
create_category 🟡 name, group_id
update_category 🟡 id, name?, group_id?, hidden?
delete_category 🔴 id, confirm
list_transactions 🟢 account_id, start_date?, end_date?
create_transaction 🟡 account_id, date, amount, payee_name?, category?, notes?
update_transaction 🟡 id, date?, amount?, payee_name?, category?, notes?, cleared?
delete_transaction 🔴 id, confirm
list_accounts 🟢
get_account_balance 🟢 account_id
create_account 🟡 name, offbudget?
update_account 🟡 id, name?, offbudget?
list_budget_months 🟢
get_budget_month 🟢 month (YYYY-MM)
set_budget_amount 🟡 month, category_id, amount
set_budget_carryover 🟡 month, category_id, enabled
list_schedules 🟢
create_schedule 🟡 account_id, amount, start_date, name?, frequency?, category_id?, payee_name?
update_schedule 🟡 id, + any schedule field
delete_schedule 🔴 id, confirm
list_payees 🟢
create_payee 🟡 name
update_payee 🟡 id, name?, category?
delete_payee 🔴 id, confirm

Troubleshooting

  • "actual-mcp is not configured"ACTUAL_SERVER_URL or ACTUAL_SYNC_ID is missing from the client's env block.
  • Can't connect / auth fails — check the URL is the sync server (not the desktop app) and the password matches. Test it by logging into that URL in a browser.
  • Wrong or empty data — the Sync ID points at a different budget file. Re-copy it from Settings → Show advanced settings → Sync ID.
  • Tools don't appear in Claude Desktop — fully quit and reopen; check the config file is valid JSON.

Verify your setup

Before wiring it into an agent, confirm the server can reach your Actual instance. This prints only counts — never names or amounts:

ACTUAL_SERVER_URL=https://actual.example.com \
ACTUAL_SYNC_ID=your-sync-id \
ACTUAL_PASSWORD=your-password \
npm run smoke
✓ connected to Actual
  accounts:        1
  categories:      63
  category groups: 17
  budget months:   21
✓ clean shutdown

Develop

npm install
npm run build
npm run inspect          # @modelcontextprotocol/inspector against the built server
npm run smoke            # connection check (counts only)

License

MIT © Mauricio Juba · TYPE:ZERO

Not affiliated with Actual Budget. Built on the official @actual-app/api.

Recommended Servers

playwright-mcp

playwright-mcp

A Model Context Protocol server that enables LLMs to interact with web pages through structured accessibility snapshots without requiring vision models or screenshots.

Official
Featured
TypeScript
Magic Component Platform (MCP)

Magic Component Platform (MCP)

An AI-powered tool that generates modern UI components from natural language descriptions, integrating with popular IDEs to streamline UI development workflow.

Official
Featured
Local
TypeScript
Audiense Insights MCP Server

Audiense Insights MCP Server

Enables interaction with Audiense Insights accounts via the Model Context Protocol, facilitating the extraction and analysis of marketing insights and audience data including demographics, behavior, and influencer engagement.

Official
Featured
Local
TypeScript
VeyraX MCP

VeyraX MCP

Single MCP tool to connect all your favorite tools: Gmail, Calendar and 40 more.

Official
Featured
Local
graphlit-mcp-server

graphlit-mcp-server

The Model Context Protocol (MCP) Server enables integration between MCP clients and the Graphlit service. Ingest anything from Slack to Gmail to podcast feeds, in addition to web crawling, into a Graphlit project - and then retrieve relevant contents from the MCP client.

Official
Featured
TypeScript
Kagi MCP Server

Kagi MCP Server

An MCP server that integrates Kagi search capabilities with Claude AI, enabling Claude to perform real-time web searches when answering questions that require up-to-date information.

Official
Featured
Python
E2B

E2B

Using MCP to run code via e2b.

Official
Featured
Neon Database

Neon Database

MCP server for interacting with Neon Management API and databases

Official
Featured
Exa Search

Exa Search

A Model Context Protocol (MCP) server lets AI assistants like Claude use the Exa AI Search API for web searches. This setup allows AI models to get real-time web information in a safe and controlled way.

Official
Featured
Qdrant Server

Qdrant Server

This repository is an example of how to create a MCP server for Qdrant, a vector search engine.

Official
Featured