bookie
An MCP server that keeps books for freelancers and landlords — driven from Claude or GPT instead of QuickBooks.
Ask your LLM to import a bank statement, categorize spending, reconcile a month, or generate a Schedule C. Bookie provides the correct double-entry ledger underneath — so the model reasons over real numbers, not a spreadsheet it's improvising on the fly.
This is for you if: you're a solopreneur, freelancer, or rental property owner already living in Claude or GPT, you're comfortable with a 5-minute setup, and you want books that are actually correct.
Not for you if: you want a dashboard UI, you need multi-user access, or you're satisfied with QuickBooks / a spreadsheet.
What you need before starting
- Node ≥ 24
- neonctl —
npm install -g neonctl(free Neon account; needed for the DB) - An MCP-capable host: Claude Desktop, Claude.ai, Cursor, VS Code, or any host supporting the MCP stdio or HTTP transport
Quick start (local, stdio)
Recommended — from source, fully automated:
git clone https://github.com/yuens1002/bookie
cd bookie
npm install
npm run setup # creates Neon DB, generates secrets, writes .env, runs db:push
npm run build
npm run setup opens a browser to log into Neon; new to Neon? Use the "Sign up for an account" link and pick GitHub/Google/Microsoft rather than email+password — it completes in the same browser round-trip, no email-verification detour that could interrupt the CLI mid-wait.
npm run setup prints a ready-to-paste Claude Desktop config block at the end:
{
"mcpServers": {
"bookie": {
"command": "node",
"args": ["/absolute/path/to/bookie/dist/index.js"],
"env": {
"BOOKIE_DB_URL": "<printed by setup>",
"BOOKIE_DB_DIRECT_URL": "<printed by setup>",
"BOOKIE_API_KEY": "<printed by setup>"
}
}
}
}
Adding another machine to the same ledger — no clone needed: stdio is a local process — every machine running a stdio MCP client spawns its own copy of the server, so each one otherwise needs its own checkout. Once the Neon DB is provisioned (via npm run setup above, on any one machine), every other machine just needs the same connection strings — no git clone, no npm install, no npm run build to keep in sync. Point that machine's MCP host at the published package instead:
{
"mcpServers": {
"bookie": {
"command": "npx",
"args": ["-y", "bookie-mcp"],
"env": {
"BOOKIE_DB_URL": "<same value as your first machine>",
"BOOKIE_DB_DIRECT_URL": "<same value as your first machine>",
"BOOKIE_API_KEY": "<same value as your first machine>"
}
}
}
}
npx fetches and runs the published version on demand — every machine pointed at the same BOOKIE_DB_URL shares one ledger, without any of them (besides the original) needing a checkout.
Bootstrapping a DB without cloning at all: if you don't have connection strings yet from any machine (e.g. you created the Neon project manually instead of via npm run setup), push bookie's bundled schema directly:
mkdir bookie-mcp && cd bookie-mcp
npm install bookie-mcp
BOOKIE_DB_URL=<pooled> BOOKIE_DB_DIRECT_URL=<direct> npx prisma db push --schema=node_modules/bookie-mcp/prisma/schema.prisma
Then use the same npx -y bookie-mcp config above.
Quick start (remote, HTTP — for Claude.ai mobile)
Deploy to Railway with one click:
Railway pulls the pre-built image from GHCR — no source build needed. Set the required env vars when prompted. See docs/DEPLOYING.md for the full walkthrough (env var reference, Neon + Resend setup, Claude.ai OAuth connector).
Why no UI?
The host LLM already reads CSVs, sees receipt images, and writes prose. Bookie owns the things an LLM shouldn't improvise: a correct double-entry ledger, integer-cent money math, deterministic categorization rules, and reproducible reports. The model handles language and vision; the server handles the books.
Tools
The full, always-current tool reference lives in docs/TOOLS.md (regenerate with npm run docs:tools). Today:
| Tool | What it does |
|---|---|
manage_accounts | Create/list/archive accounts (segment-scoped categories carry a tax line) |
add_transaction | Record one balanced double-entry (money flows from → to) |
split_transaction | One payment leg + N category legs (a receipt split across categories) |
import_transactions | Import a bank/card CSV as balanced entries — preview → confirm, with dedup |
manage_rules | Create/list/delete/test/suggest auto-categorization rules (categorize → account/property, or exclude) that power import-preview suggestions; action=suggest scans past categorizations and returns candidate rules for descriptions with 2+ occurrences |
categorize_transaction | Re-categorize the income/expense leg of an existing entry — explicit account or apply a stored rule |
reconcile | Match a bank/card statement CSV against the ledger and mark postings cleared — preview then commit |
manage_receipts | Attach, list, delete, or get a signed download URL for receipt data; optionally upload the original file (JPEG, PNG, WEBP, HEIC, or PDF) to Railway Bucket storage |
generate_report | Monthly reconciliation summary, or fiscal-year Schedule C / Schedule E tax P&L |
export_report | Render any report as markdown or CSV |
send_report | Run a report and email it via Resend |
query_transactions | List entries + postings by date range / account |
account_balances | Current balance per account |
Resources
Bookie exposes two MCP resources that an LLM can read without calling a tool:
| Resource URI | MIME type | What it contains |
|---|---|---|
bookie://accounts | application/json | All active accounts with their current balances |
bookie://reports/{year} | text/markdown | Annual fiscal snapshot: Schedule C, Schedule E, and a one-row-per-month summary (opening balance, net income, cleared postings count) |
Prompts
Three canned workflow prompts guide the LLM through common bookkeeping tasks:
| Prompt | Parameters | Purpose |
|---|---|---|
monthly-close | year, month | Step-by-step month-end close: import CSV → categorize → reconcile → report → (optional) email |
categorize-uncategorized | (none) | Find journal entries with no income/expense leg and walk through categorizing each |
prepare-tax-summary | year | Generate Schedule C + E, export as markdown and CSV, optionally email |
Configuration
See .env.example for the full reference. Key variables:
| Variable | Purpose |
|---|---|
BOOKIE_TRANSPORT | stdio (default) or http |
BOOKIE_DB_URL | Neon pooled connection string |
BOOKIE_DB_DIRECT_URL | Neon direct connection string (for db push) |
BOOKIE_API_KEY | Static Bearer token (Claude Desktop / direct API) |
PUBLIC_URL | Public HTTPS base URL of the deployed server (Claude.ai connector) |
JWT_SECRET | HS256 signing secret for OAuth JWT access tokens |
OAUTH_CLIENT_ID | OAuth client ID (default: claude-ai-connector) |
OAUTH_CLIENT_SECRET | Required when using OAuth: /authorize refuses all requests when unset (prevents any visitor from authorizing); /token also validates it. Enter this value in the Claude.ai connector settings. |
RESEND_API_KEY | Resend API key for send_report |
RESEND_FROM | Verified sender address for send_report (e.g. Bookie <reports@yourdomain.com>) |
AWS_ENDPOINT_URL / AWS_S3_BUCKET_NAME / AWS_ACCESS_KEY_ID / AWS_SECRET_ACCESS_KEY / AWS_DEFAULT_REGION | Railway Bucket credentials — auto-injected when you connect a bucket to the service (use AWS SDK Generic style); enables receipt file upload in manage_receipts |
Docs
- Architecture — data model, transports, layering
- Deploying — Railway + Neon + Resend setup
- Roadmap — phased plan
- Changelog — what shipped
- Releasing — versioning + release process
- Tools — generated manual
- Contributing — setup, conventions, tests
Safety
Bookie stores financial data in your Neon Postgres database; connection strings live in .env (gitignored) and Railway env vars — never commit them. The HTTP transport requires auth on every /mcp request: either a static Bearer token (BOOKIE_API_KEY) or an OAuth JWT issued by the /token endpoint. Always set at least one before exposing the server beyond localhost. See docs/DEPLOYING.md for the full Claude.ai connector OAuth setup.
License
MIT