Open Accountant

Your AI bookkeeper. Follow the money.
Privacy-first financial assistant for your terminal.

Import bank statements, categorize transactions with AI, surface spending anomalies, and get actionable advice — all without your financial data leaving your machine.

Named after Frank J. Wilson, the forensic accountant who followed the money to convict Al Capone.

• Bank CSV import — Auto-detects Chase, American Express, and generic CSV formats
• Monarch Money sync — Pull transactions directly from your Monarch Money account
• AI categorization — Classifies transactions into 18 spending categories using your choice of LLM
• Spending summaries — Breakdowns by category, merchant, or time period with comparisons
• Anomaly detection — Flags duplicate charges, unusual spikes, and forgotten subscriptions
• Export — Save filtered transactions to CSV or XLSX
• Web search — Research merchants and financial questions (Exa, Perplexity, Tavily, Brave)
• MCP extensibility — Add tools via Model Context Protocol servers
• Orchestration — Chain tools sequentially or run parallel teams for complex workflows
• Skills — Multi-step workflows like subscription audits, extensible with custom skills
• Semantic transaction search — The dashboard Transactions search understands meaning, not just substrings: a query with no exact matches falls back to on-device embedding search (wilson --index builds the vectors) with ranked, scored results — nothing but the one-time model download ever leaves the machine
• 9 LLM providers — OpenAI, Anthropic, Google, xAI, Moonshot, DeepSeek, OpenRouter, LiteLLM, and Ollama (local)
• Offline dashboard transactions — the dashboard keeps a sync-fed local mirror (wa-sqlite on OPFS, inlined in the single-file build); when the server is unreachable you can still browse, search, and filter transactions with the same results, and entity assignment shows an explicit "requires connection" state instead of failing silently
• Offline overview cards — the heatmap, streak, weekly summary, budget countdown, savings sparkline, category donut, P&L, and budget bars all recompute locally from the mirror (same SQL and math as the server); cards whose engine is server-side (alerts, net worth, cash forecast) say so explicitly instead of failing

The React dashboard stores a read-only mirror of the active profile's transactions, entities, budgets, and categories in the browser (wa-sqlite on OPFS, keyed per profile). While the server is reachable the mirror stays fresh via a periodic full pull; when the server is unreachable, the transactions tab serves browsing, search, and filtering from the last-synced mirror — the same results the server would return for the same query — and the overview tab renders its eight approved cards from locally recomputed aggregations (the server's own SQL constants and calendar/rollup math are shared code, so both sides agree by construction). Cards that depend on server-only state — the alerts engine, net worth/accounts, and the cash forecast — render an explicit "Unavailable offline" note rather than misleading zeros. Entity assignment (and every other write) stays online-only: offline it shows a "requires connection" state rather than failing silently, and nothing is queued.

Privacy note: the browser mirror is unencrypted at rest (browser origin storage holds plaintext, unlike the SQLCipher-encrypted CLI database). It mirrors exactly what the dashboard already renders in that same browser, it can be dropped and re-synced at any time, and full-disk encryption (e.g. FileVault) protects it at the OS layer.

• Bun v1.1+
• An LLM provider: either Ollama running locally or an API key for a cloud provider

git clone https://github.com/openaccountant/wilson.git
cd open-accountant
bun install
cp env.example .env   # edit with your API key(s)
bun start

1. Import transactions — Drop a CSV into the chat: Import my transactions from ~/Downloads/chase.csv
2. Categorize — Categorize my uncategorized transactions
3. Explore — What did I spend on dining last month?
4. Audit — Find any unusual charges or forgotten subscriptions

Switch providers at any time with /model.

Type exit or quit to close. Press Esc to cancel a running operation.

Copy env.example to .env and set at least one provider API key. See the file for all options including web search keys and Ollama base URL.

Add external tool servers in ~/.openaccountant/mcp.json:

{
  "servers": {
    "my-server": {
      "command": "npx",
      "args": ["-y", "my-mcp-server"],
      "env": {}
    }
  }
}

MCP tools appear automatically in Open Accountant's tool registry.

Drop a folder with a SKILL.md file into any of these directories:

Skills defined later in this list override earlier ones with the same name.

src/
  agent/          # Core agent loop, tool execution, context management
  components/     # TUI components (chat log, editor, prompts)
  controllers/    # Agent runner, model selection, input history
  db/             # SQLite schema, queries, database init
  mcp/            # MCP client, adapter, config
  model/          # LLM abstraction and provider implementations
  orchestration/  # Chains (sequential) and Teams (parallel) workflows
  skills/         # Skill discovery, loading, and built-in skills
  tools/          # All tool implementations
    categorize/   # AI transaction categorization
    export/       # CSV/XLSX export
    import/       # CSV import with bank parsers, Monarch sync
    query/        # Transaction search, spending summary, anomaly detection
    search/       # Web search providers (Exa, Perplexity, Tavily, Brave)
  utils/          # Shared utilities

All data is stored locally in ~/.openaccountant/data.db (SQLite).

Contributions are welcome! Please open an issue first to discuss what you'd like to change.

See the issue templates for bug reports and feature requests.

MIT