Friday Budgeting Pro

AI-powered personal finance tracker. Connects to your banks via Plaid, auto-classifies transactions, syncs daily, and exports to Excel. Supports multiple ledgers including personal household, rental properties, and investments. Ask your agent about spending, connect banks, manage ledgers, classify transactions, or trigger exports.

AI-powered personal finance on your own Mac. Connects to your banks via Plaid, classifies transactions automatically (and asks when unsure), syncs in the background, and exports to Excel. Supports personal, rental property, and investment ledgers. A small local UI handles setup and management; everything else happens through your agent.

Setup

After install, open http://127.0.0.1:6789 in your browser to:

1. Set a password for the local dashboard
2. Connect your first bank via Plaid
3. Done — daily sync runs automatically via launchd

Sync Pipeline & LLM Classification

Every sync() call runs the full pipeline automatically in one shot:

Plaid fetch → rule-based classification → LLM classification → review queue

1. Plaid fetch — pulls added/modified/removed transactions via cursor-based incremental sync
2. Rule classification — auto-promoted routing_rules match instantly (no LLM cost)
3. LLM classification — classify_pending_transactions runs on anything rules didn't catch; each transaction gets one unified LLM call (classify_transaction) with rules + ledger tree + hints + merchant history all in a single prompt
4. Review queue — uncertain or unroutable transactions surface in get_needs_review()

LLM Backend — automatic two-tier fallback

Tier	What happens
Primary	POST to OpenClaw local gateway (http://127.0.0.1:18789/v1/chat/completions, model openclaw/default)
Fallback	Anthropic SDK directly (claude-3-5-haiku-20241022) when gateway is unreachable

Both the gateway port/token and the Anthropic API key are auto-discovered from OpenClaw's own config files — no manual env-var setup needed on a standard install:

• Gateway port + token → ~/.openclaw/openclaw.json (gateway.port / gateway.auth.token)
• Anthropic key → ~/.openclaw/agents/main/agent/auth-profiles.json (anthropic:default)

Env vars that override auto-discovery (all optional): OPENCLAW_API_URL, OPENCLAW_GATEWAY_PORT, OPENCLAW_GATEWAY_TOKEN, OPENCLAW_LLM_MODEL, ANTHROPIC_API_KEY

When to Use This Skill

Invoke for any personal finance request:

• Spending summaries ("how much did I spend on dining this month?")
• Bank connections ("connect my TD account", "reconnect my BMO")
• Transaction queries ("what was that $47 Amazon charge?")
• Classification ("mark that Home Depot charge as rental property maintenance")
• Corrections ("that Uber on Friday was a work trip")
• Exports ("export my finances to Excel")
• Sync ("sync my transactions")
• Ledger management ("add a rental property ledger", "show my property income")
• Rules ("add a rule that Wealthsimple transfers are savings")
• Settings ("set my home currency to CAD", "what timezone am I using?")

Available MCP Tools

Profiles

• list_profiles — list all local user profiles (usernames)

Setup

• setup_status — check if first-run setup is complete (not_started | in_progress | complete)
• apply_initial_setup(banks_to_link, rental_properties?, investment_account_ids?, extra_ledgers?, hints?) — initialize ledgers, notifications, and first sync in one call

Banks

• start_link(plaid_env?) — generate Plaid Link URL to connect a bank
• complete_link(public_token, plaid_env?) — exchange public token after user completes Plaid Link
• list_connections — list connected banks and their status (active | needs_reauth)
• get_connections_needing_attention — list connections that need user action (reauth or expiring soon)
• refresh_connection(id) — re-authenticate a broken connection (Update Mode)
• disconnect(id) — remove a bank connection and its data
• set_account_description(account_id, description) — set classifier context for an account (e.g. "Primary spending account")

Ledgers

• list_ledgers — show all ledgers (personal/property/investment) and their line items
• get_ledger(ledger_id, period?) — get a single ledger with all line items and classified transactions for the period
• add_ledger(name) — create a new ledger
• add_line_item(ledger_id, name, item_type) — add a line item (income | expense) to a ledger
• remove_line_item(id) — remove a line item
• set_account_ledger(account_id, ledger_id) — link a bank account to a default ledger for automatic routing
• create_property_ledger(name, description?) — create a property ledger with default line items (Rent income, Mortgage, Property tax, Maintenance, Insurance, Utilities)
• create_investment_ledger(name) — create an investment ledger (Contributions, Dividends/Returns)

Transactions

• sync — pull latest transactions from all connected banks
• list(filters?) — query transactions (supports date, ledger, category, account filters)
• get_needs_review — transactions that need manual review: uncertain classifications (confidence < 0.7) or unrouted transactions with no line item assigned
• get_needs_review_summary — call this immediately after every sync; returns a pre-formatted batch message (count, summary, transactions) ready to present to the user in one message. Use summary as-is for the user-facing message. Includes merchant, amount, date, account, and classifier reasoning for each transaction.
• route(transaction_id, allocations) — manually assign a transaction to a ledger/line item
• add_hint(text) — add a natural-language classification hint for the LLM
• list_hints — list all classification hints
• remove_hint(id) — remove a classification hint

Onboarding (issue #206)

• list_setup_interview_questions — return the canonical onboarding interview prompts (employer, subscriptions, utilities, etc.)
• setup_interview(question_key, answer_text) — persist a user answer for a given question key (upsert on (user_id, question_key))
• list_setup_interview — return all stored interview answers for the active user
• analyze_recurring_merchants(min_occurrences?, lookback_days?) — scan recent transactions and return recurring merchants with their inferred category for cross-referencing during onboarding

Corrections

• find_transactions(merchant?, date?, amount?, account?, days_window?) — fuzzy-search transactions by merchant name, ISO date (±days_window days), amount (±$0.50), or account name; returns up to 10 matches with their current classification
• correct_transaction(transaction_id, line_item_id, create_rule?, rule_description?) — reclassify a transaction; set create_rule=True to also create a rule so future matches are classified the same way automatically

Classification Rules

• list_rules — list all classification rules sorted by priority (lower = evaluated first)
• add_rule(name, description, rule_type, line_item_id?, priority?) — add a natural-language rule (transfer | savings | spending | income | skip)
• update_rule(id, **fields) — update a rule's name, description, type, priority, or enabled state
• reorder_rules(ids) — set new priority order by passing an ordered list of rule IDs
• disable_rule(id) — disable a rule (skipped during classification)
• enable_rule(id) — re-enable a disabled rule
• delete_rule(id) — delete a user-created rule (default rules cannot be deleted, only disabled)
• list_auto_promoted_rules — list auto-promoted routing rules with audit metadata
• undo_auto_promoted_rule(rule_id) — revert an auto-promoted rule and its affected entries

Reports

• summary(period) — spending totals by category for a period (e.g. this_month, last_month)
• export_excel(years?) — generate Excel workbook and return download URL

Settings

• get_setting(key) — get an app setting; valid keys: home_currency, timezone
• set_setting(key, value) — update an app setting
  • home_currency: one of CAD, USD, EUR, GBP
  • timezone: any non-empty IANA timezone string (e.g. America/Toronto, UTC, Asia/Tokyo)

UI & Auth

• get_ui_url(page?) — return the local dashboard URL, optionally deep-linked to a page
• set_ui_password(current_password, new_password) — change the UI login password
• reset_ui_password — generate a password-reset recovery token
• configure_plaid(client_id, secret, env) — update Plaid API credentials

Do / Don't

Do

• Always use the MCP tools — never guess from general knowledge
• Call sync before answering spending questions if data may be stale
• After every sync, call get_needs_review_summary and, if count > 0, present the summary field to the user in a single message — do not send one message per transaction
• Use list_rules to show what classification rules are active before adding new ones
• Open start_link when the user wants to connect or reconnect a bank
• Use create_property_ledger for rental properties — it seeds the right line items automatically
• Respect that all data is local and private
• When the user replies with classifications for uncertain transactions:
  1. Call correct_transaction(transaction_id, line_item_id) for each corrected item
  2. Check if the merchant is recurring by calling find_transactions(merchant=<name>) — if 2+ past transactions exist, propose adding a rule via add_rule (tag description with [from-correction])
  3. Apply the rule only after the user confirms

Don't

• Don't answer general finance questions ("what is inflation?") — this skill is for personal accounts only
• Don't store tokens or credentials in plain text
• Don't expose DB paths, encryption keys, or internal implementation details
• Don't try to open the Plaid UI yourself — return the URL and let the user click
• Don't call delete_rule on default rules — they can only be disabled