Scheduled money reviews your agent runs by itself

The three prompts

Each prompt ends by telling the agent that nobody can answer questions, because a scheduled run has nobody at the keyboard. Save them as they are or edit the wording. The commands below use the weekly prompt, and the other two go in the same place.

Weekly review, Monday morning

Using BankBridge, review the last 7 full days across my accounts. Report total spent, total received, the five largest charges, and spending by category next to the 7 days before. Leave out card payments and transfers between my own accounts. Put any notice about a bank that needs reconnecting first. Keep it under 15 lines. Nobody can answer questions during this run, so state any assumption and finish.

The result looks like this. These are demo numbers from BankBridge's fictional sample accounts, for the week of October 27, 2026:

New-subscription alert, daily or weekly

Using BankBridge, list every charge from the last 7 days. For each merchant, check get_merchant_history over the 12 months before that. Report merchants with no earlier charge: name, amount, date and last four digits of the account. Mark the ones that look like a subscription or trial. If there are none, say so in one line. Nobody can answer questions during this run, so state any assumption and finish.

This catches a first charge, which a recurring-charge search can't see until the second one posts. The new-subscription alert guide adds a version that runs with no model at all.

Month-end report, on the 2nd

Using BankBridge, build a report for last calendar month. Show income, spending and net, leaving out card payments and transfers between my own accounts. Compare each category with the month before, list the ten largest merchants, and list recurring charges that changed price or are new. Put any notice about a bank that needs reconnecting first. Nobody can answer questions during this run, so state any assumption and finish.

The exclusions matter. In the demo accounts, the raw get_monthly_cashflow totals for October 2026 are $6,142.18 in and $5,973.86 out. Those include $1,412.63 of card payments and a $500 move from checking to savings. Without them, the same month is $5,642.18 in and $4,061.23 out.

Running a business? The LLC month-end close uses a different prompt, with owner draws and a CSV for your accountant.

Where to run them

Point Hermes Agent or OpenClaw at a model on your own hardware and the conversation stays there. BankBridge and the bank-data provider still handle the numbers on the way to your machine, and BankBridge keeps no copy. The commands below come from each project's documentation, read on November 6, 2026. These tools change quickly, so check the linked page if a flag is rejected.

Claude scheduled tasks

Anthropic's scheduled tasks page says a task can use "any connectors and plugins you've set up in Cowork," and that tasks run remotely, so they keep their schedule while your computer is asleep. It lists them on Pro, Max, Ultra, Team and Enterprise. BankBridge is added as a connector, so that's the documented path. We're going by those docs and haven't run this path ourselves.

1. Add BankBridge from Claude's connector directory and connect a bank.
2. Describe the task and its timing in a conversation, or open Scheduled in the left sidebar, click New task and choose Set up manually.
3. Paste a prompt from above. The manual form offers hourly, daily, weekly, weekdays or manual.
4. Run the task on demand once and read the result.

Anthropic's page doesn't list a monthly option, so these suit the weekly review and the alert best. For the month-end report, run the saved task on demand at the start of the month, or use a Claude Code cloud task with a custom schedule. On Team and Enterprise, admins control Cowork access, and an Owner adds the connector first: see the Team and Enterprise setup.

Claude Code

Claude Code has three schedulers. Two fit this job.

Cloud task. Runs with your laptop closed. Create it at claude.ai/code/routines, or run this in any session:

/schedule every Monday at 8am, review the last 7 days of my bank accounts with BankBridge

Paste the full prompt when Claude asks what the task should do. The docs say a cloud task can use connectors on your claude.ai account, and all of them are included by default, so remove the ones this task doesn't need. A server you added with claude mcp add lives on your machine and won't appear, so add BankBridge at claude.ai/customize/connectors. For a monthly run, pick Custom in the task's edit dialog and enter a cron expression such as 0 13 2 * *, which the docs read as UTC. Cloud tasks are in research preview.

Desktop task. Runs on your machine while the desktop app is open and the computer is awake. A run can stall on a permission prompt, so click Run now after creating it and choose always allow for each BankBridge tool.

Skip /loop here: it needs an open session and recurring loops expire after seven days. To run a plugin command by hand instead, /bankbridge:weekly-review and /bankbridge:monthly-report produce similar reports.

Hermes Agent

Add the server to ~/.hermes/config.yaml, as in Hermes's MCP docs. Your key is in BankBridge settings.

mcp_servers:
  bankbridge:
    url: "https://bankbridge.money/api/mcp"
    headers:
      Authorization: "Bearer bbk_your_key"

Then create the job with the cron command:

hermes cron create "0 8 * * 1" \
  "Using BankBridge, review the last 7 full days across my accounts. ..." \
  --name "Weekly money review" \
  --timezone America/Chicago \
  --deliver telegram

• Jobs only run while the gateway does. Install it as a service with hermes gateway install, and check it with hermes cron status.
• A job gets the tools set for the cron platform in hermes tools, so check that BankBridge's toolset, mcp-bankbridge, is included.
• Leave out --deliver from the command line and the report saves to ~/.hermes/cron/output/.
• Test with hermes cron run <job_id>. Hermes's docs say tools marked read-only don't prompt, which matters because nobody is there to approve them.

OpenClaw

Add the server with the command on our OpenClaw page, then create an automation. openclaw cron is an alias for the same commands.

openclaw automations create "0 8 * * 1" \
  "Using BankBridge, review the last 7 full days across my accounts. ..." \
  --name "Weekly money review" \
  --tz America/Chicago \
  --session isolated \
  --announce \
  --channel telegram \
  --to "123456789"

• Automations run inside the Gateway, so it has to be running when a job comes due.
• Use --no-deliver in place of the three delivery flags to keep the output internal.
• Test with openclaw automations run <job-id>, then read the result with openclaw automations runs <job-id>.

Plain cron

Cron needs an agent with a non-interactive mode. This one uses claude -p. First, put the server config and the prompt in files, so your key stays out of the crontab:

mkdir -p ~/.bankbridge
cat > ~/.bankbridge/mcp.json <<'EOF'
{"mcpServers":{"bankbridge":{
  "type":"http",
  "url":"https://bankbridge.money/api/mcp",
  "headers":{"Authorization":"Bearer bbk_your_key"}
}}}
EOF
chmod 600 ~/.bankbridge/mcp.json
# Save the weekly prompt to ~/.bankbridge/weekly.txt

Generate a login token with claude setup-token and add a small script at ~/.bankbridge/weekly.sh, with its mode set to 700:

#!/bin/sh
export CLAUDE_CODE_OAUTH_TOKEN="your_token"
exec "$HOME/.local/bin/claude" -p "$(cat "$HOME/.bankbridge/weekly.txt")" \
  --strict-mcp-config \
  --mcp-config "$HOME/.bankbridge/mcp.json" \
  --allowedTools "mcp__bankbridge__*" \
  < /dev/null

Point the path at wherever which claude says it lives, because cron starts with a short PATH. Then crontab -e:

MAILTO=you@example.com
0 8 * * 1 $HOME/.bankbridge/weekly.sh

We ran this command with BankBridge's demo key, bbk_demo, which returns fictional accounts. Use it for your first test. --strict-mcp-config loads only the servers in that file, and --allowedTools approves only BankBridge's tools, because an unattended run can't answer a prompt.

Cron emails whatever the job prints when MAILTO is set and the machine can send mail. Otherwise append >> $HOME/.bankbridge/reports.log. A laptop asleep at 8am misses that run, and cron doesn't make it up. A small always-on server avoids that.

Check the first run

A green status isn't proof. Anthropic's docs say a cloud run marked Succeeded can still have had missing connector tools, with the details in the transcript. Read the first report and check three things:

• The numbers match a transaction or two you can see in your banking app. Balances and new charges can lag at the bank, so a charge from this morning may not be there yet.
• Every bank you expect appears, and the report shows any reconnect notice at the top.
• The report landed where you'll see it, and only people who should see those numbers can read it.

Then let it run for a month before you stop checking by hand. Connect a bank to start, or try the commands with the demo data.