Run an Actual Budget ledger from your AI agent, with phone approvals for every change on paired machines.
Personal finance data is the last thing most people want an agent writing to unsupervised. The Arc MCP server is built around that worry. It ships inside the arc CLI, which sits on top of an Actual Budget ledger (either one managed by arc or your own Actual server), and arc mcp exposes the same operations the command line has, one tool per command, with identical argument names and validation.
The catalogue is wide, roughly 120 operations, grouped the way a budget is organised:
Thirteen dashboard recipes also arrive as MCP prompts (recipe-spending, recipe-budget and others), which tell an agent how to assemble an HTML dashboard from the data.
You need an arc account. The arc app, under Settings > AI agents, shows a one-line installer that carries your instance URL and credentials; that payload is a secret and grants full access to the budget it was made for. Self-hosted Actual servers get a free arc licence for one host, valid 30 days on the free plan and renewed automatically, and they must use HTTPS unless they sit on localhost or a private network. After installing, register arc mcp with your client. The installer adds a Claude Desktop entry by itself.
Pairing a machine changes the risk picture considerably. Under the default Standard preset reads run immediately, while every write waits for Face ID approval on your phone. Over MCP a gated tool waits about 45 seconds, then returns pending_approval; the agent finishes it later with arc_agent_request_status. No tool can approve its own request. arc activity keeps a hash-chained timeline of what each named agent did.
The installer is macOS-first; on Linux the CLI works but the Claude Desktop step does nothing. Unpaired payload installs run every operation without asking. Managed servers sleep when idle, so the first call can take a few seconds. A remote mode (arc mcp --http) exists for web clients behind a tunnel, and anyone holding that URL and token gets read and write access. For other finance-adjacent tools, browse our other integrations, or see the full list in that category for comparable connectors.
Check the linked publisher documentation and your connected client's tool list. Names and availability can vary with the server version and account permissions.
| Tool | What it does |
|---|---|
| arc_agent_permissions | Reports what the current agent may do on this machine: which reads, writes and deletes run, ask first or are refused. |
| arc_agent_request_status | Finishes a call that returned pending_approval, running the approved call exactly once. |
{
"mcpServers": {
"arc": {
"command": "arc",
"args": [
"mcp"
]
}
}
}Install the arc CLI first using the command from Settings > AI agents in the arc app. The installer normally writes this entry for you.
| Variable | Description | Required |
|---|---|---|
| ACTUAL_CUSTOM_HEADERS | JSON list of reverse-proxy headers for your own Actual server, overriding the stored ones for one run. | Optional |
| ARC_CLI_LICENSE | Overrides the stored arc licence used with a self-hosted Actual server. | Optional |
| ARC_AGENT | Names the agent making calls, so approvals and permissions can be set per agent. | Optional |
| ARC_APPROVAL_WAIT_SECONDS | How long a non-terminal CLI call waits for approval before exiting (default 90 seconds). | Optional |
An Odoo module that serves a read-only MCP endpoint at /mcp, scoped to each user's own access rights.
Drives DaVinci Resolve 21.1, free edition included, from Claude Code through a Lua bridge running inside the app.
Open-source screen recorder and editor whose MCP server lets an agent record, edit and export a demo video.
Ask an assistant for today's internet-wide cyber threat band, score, critical advisories and data-source health.
Lets coding agents run dev servers and workers as real systemd, launchd or Windows services, and read their logs.
A desktop downloader and toolbox that coding agents can drive locally: queue media, read files, check ports.