Spendglass

Spendglass keeps a copy of your bank transactions on your own computer, where your local AI agents can read it. It syncs your accounts through Redbark, an open-banking service, into a SQLite file. It works out who your merchants are. You see the result in a web app behind a passkey or a password, with a spending page. A read-only MCP server lets the agents on your computer answer questions about your money.

It only answers on the computer it runs on. Nothing listens beyond loopback, and there's no remote mode to get wrong.

The promise, in plain English

• Your data stays on your computer unless you set something up to send it. The server listens on 127.0.0.1 only, and there's no cloud part, no telemetry and no account with us. With an Anthropic key, merchant identification sends merchant details to Anthropic. What the agent tools answer goes to your agent's own model, and a backup mirror copies the store to the folder you choose. SECURITY.md lists what goes where.

• Nothing here can move money. The sync only reads, because its bank client makes nothing but GET requests. The agent tools only read too, and a test pins the exact list of them, so a tool that writes can't appear by accident. The web app can't change what the bank sent. It writes your labels, such as merchant names, categories and themes, along with its own settings, provider keys and sign-in records.

• Only the sync process talks to your bank. sync.py is the only code that uses the Redbark key, and it runs as a separate process that exits when it's done. The web app and the MCP server never call the bank, and SECURITY.md says where that split stops.

• It works without an AI key. Every deterministic feature runs with no model, including sync, transfers, recurring detection, trends, themes and backups. An Anthropic key adds merchant identification, and leaving it out never breaks anything.

Related MCP server: ZenMoney MCP

What you get

You get sync, the store and the web app. You also get a merchant-identity pipeline, with a lookup agent that searches the web and a review queue where you have the final say. There's an internal-transfer matcher, themes that cut across categories, a subscription detector and eight trend views. Sync runs in the background on a schedule, backups happen by themselves, and the spending page drills down to the transactions behind each number. Every chart comes with a plain-English "what am I looking at?" explainer.

Requirements

• Python 3.12 or newer.

• A Redbark API key, which starts rbk_live_…. Redbark is an Australian open-banking aggregator that works under the Consumer Data Right, or CDR. Your bank consents live in its dashboard, and Spendglass only ever reads through it. The data is in Australian dollars and shaped by the CDR, with 16 fixed bank categories and consents that last 12 months.

• Optionally, an Anthropic API key for merchant identification, set as ANTHROPIC_API_KEY in .env. The admin panel can save it there for you, after checking it with Anthropic.

Quick start

git clone https://github.com/shawn-durrani/spendglass.git
cd spendglass
./start.sh

start.sh creates .venv, installs the dependencies when they change, and won't start a second copy on a port that's already taken. On the first run it creates .env from the example. It tightens .env to 0600 on every run, and it serves the web app at http://127.0.0.1:8903.

Add your Redbark key to .env, then pull your data:

.venv/bin/python -m spendglass.sync

If you skip that step, the server's own scheduled sync picks the key up within 15 minutes.

On your first visit you set a password, using the recovery secret the terminal printed, or the one you set yourself. docs/CONFIG.md explains the three credentials and every other setting.

To keep it running unattended, install the launchd supervisor once with ops/install-supervisor.sh. It starts the app at login, restarts it after a crash and brings it back after a reboot. docs/OPERATIONS.md says how it works.

Connect your agents

claude mcp add -s user spendglass -e PYTHONPATH=<repo> -- <repo>/.venv/bin/python -m spendglass.mcp_server

That registers sixteen read-only tools. None of them writes, and a test pins the exact list, because the tool list is the security boundary. docs/MCP.md says what they cover.

Documentation

docs/README.md lists every document by what you're trying to do. The short version:

• docs/CONFIG.md: every setting, the keys and the credentials.

• docs/OPERATIONS.md: the supervisor, syncing, backups and restoring one.

• docs/MCP.md: the agent tools and what they promise.

• docs/MERCHANTS.md: how merchants get their names, what the miners cost, what counts as spend, and themes.

• ARCHITECTURE.md: the design decisions that are settled.

Security

The web app answers only on loopback, and it checks the Host header so a web page can't reach it through DNS rebinding. It sits behind a passkey or a password, keeps only hashes of your password and sessions, and refuses requests from other sites. A provider key is checked before it's saved and never shown back to you. SECURITY.md holds the threat model and how to report a problem, and ARCHITECTURE.md holds the design rules behind it.

Licence

MIT. One third-party component ships in this repository under a different licence, and ACKNOWLEDGEMENTS.md names it.