cfo-chrome-download-statements

name: cfo-chrome-download-statements description: Use when statement files are not on disk yet and the agent needs human-in-the-loop browser assistance through the user's current Chrome session to find official portals and download CSV or PDF statements.

CFO Chrome Download Statements

Use this skill when statement files are not on disk yet and the agent needs to guide a human through downloading CSV or PDF evidence from a bank, credit card, brokerage, or payment platform website.

Intent

Keep this workflow LLM-driven and human-in-the-loop:

• The LLM decides how to navigate based on the live site, prior manifests, and user input.
• The human logs in, completes MFA, confirms the correct account, confirms the date range, and confirms each ambiguous download action.
• The agent should prefer controlling the user's current Chrome session through Chrome DevTools MCP.
• The agent should not rely on deterministic site-specific scripts, brittle selector packs, or hardcoded month-selection logic.

Preferred Browser Control

Prefer Chrome DevTools MCP attached to the user's current Chrome browser.

Ask the user to enable remote debugging in:

chrome://inspect/#remote-debugging

Recommended MCP config example:

"chrome-devtools": {
  "command": "npx",
  "args": [
    "chrome-devtools-mcp@latest",
    "--autoConnect"
  ]
}

Notes:

• --autoConnect helps the MCP attach to the active Chrome session.
• Do not tell the user to launch a separate browser instance unless the current Chrome session cannot be used.

Fallback: MCPorter CLI

If the current agent environment does not expose chrome-devtools as an MCP server, fall back to mcporter.

MCPorter can discover already-configured servers, call them directly, and generate shareable CLIs, including ad-hoc stdio commands such as npx -y chrome-devtools-mcp@latest.

Useful examples:

# discover configured MCP servers
npx mcporter@latest list

# inspect available chrome-devtools tools if already configured
npx mcporter@latest list chrome-devtools

# call a chrome-devtools tool directly
npx mcporter@latest call chrome-devtools.list_pages --output markdown

# ad-hoc fallback if chrome-devtools is not already configured
npx mcporter@latest generate-cli \
  --command "npx -y chrome-devtools-mcp@latest --autoConnect" \
  --output ./tmp/chrome-devtools.ts

Use this fallback when:

• the host agent does not have a working chrome-devtools MCP integration
• you need to inspect or call the server outside the current agent runtime
• you want a temporary dedicated CLI for the statement-download workflow

Inputs

Prefer ledger-local context when present:

• capture/statement-export.yaml
• prior capture/*-export-manifest.yaml
• the ledger account map in main.beancount or accounts.beancount

Use those as context only. They are not a deterministic execution plan.

Core Rules

1. Use web search when the correct login page, statement page, or export page is unclear.
2. Restrict that search to official institution domains whenever possible.
3. Treat prior manifests as historical context, not as a deterministic rule for "next month".
4. Ask the human to confirm the exact account and date range before any download click.
5. Preserve raw filenames and archive copies before any normalization or import.
6. Never store credentials, OTP codes, security answers, or copied session data in the repo.
7. Prefer the user's current Chrome session over spawning a separate automation browser.

Workflow

1. Build context

Read the configured institutions and accounts, then summarize:

• institution
• ledger account
• preferred export types
• prior confirmed export periods if any
• likely missing periods or activity windows that should be reviewed

If prior confirmed export periods exist, propose a candidate overlap into the prior range so delayed postings are less likely to be missed. Treat the overlap size as human-confirmed context based on the institution's actual posting lag.

This summary is for the LLM and the human. It is not a deterministic instruction set.

2. Find the right portal

If capture/statement-export.yaml already names the institution, but not the exact login or export URL, use web search first.

Search policy:

• prefer official domains only
• verify the domain before opening login
• avoid SEO junk, affiliate pages, and support-forum guesses

Good examples:

• TD Canada Trust business banking login site:td.com
• Chase download account activity csv site:chase.com
• Interactive Brokers activity statements csv site:interactivebrokers.com
• Wealthsimple monthly statement download site:wealthsimple.com

3. Use Chrome DevTools MCP as a live assistant

Use Chrome DevTools MCP or the available Chrome browser tool to:

• connect to the user's current Chrome session
• open or inspect the official login page
• keep the session visible to the human
• inspect page structure after login
• help find likely statement or activity sections
• help identify export controls

Do not assume selectors will stay stable. Re-read the live DOM as needed.

4. Keep selection non-deterministic

The LLM may propose likely next exports based on prior manifests, but it must present them as:

• candidate periods
• candidate export types
• candidate account views
• candidate overlap windows versus the prior export when delayed posting risk exists

The human must confirm:

• correct account
• correct date range
• overlap size if re-exporting part of the prior window
• CSV vs PDF vs both
• whether the portal view is statements, transactions, tax slips, or something else

5. Record provenance

After each successful download, record:

• institution
• ledger account
• confirmed period
• confirmed overlap used if any
• export type
• original filename
• saved path
• manifest path

Output

• raw CSV/PDF statement files downloaded by the human with agent assistance
• an updated export manifest created by the higher-level workflow
• a clear handoff to /capture, /bank-import, or /reconcile

Anti-Patterns

• hardcoded selectors for a specific institution
• hardcoded month-stepping logic
• assuming the prior export end date is always a safe hard cutoff
• assuming the newest statement is always the correct target
• storing login secrets or browser session artifacts in the repo
• silently downloading files without human confirmation