Help · Settings and more
Use Tallybook with AI agents
Connect Claude, OpenAI's Codex and Agents SDK, OpenClaw, or any MCP-capable agent to Tallybook, so it can answer questions about your books and, if you allow it, keep them.
Tallybook can run without its window as an MCP server, so an AI agent can read your books and, if you allow it, keep them: categorize the review inbox, record cash, set a month’s budget, check what’s due.
MCP is an open standard, so this isn’t tied to one company’s AI. It works with anything that can start a local MCP server: Claude, OpenAI’s Codex and Agents SDK, OpenClaw, and most coding tools and agent frameworks.
It runs on your computer, against your book file, like the app does. Nothing is uploaded anywhere by Tallybook; what an agent does with the answers is up to the agent you connect, and to the AI service behind it.
Turn it on
The installed app is the server. Start it with --mcp:
| System | Command |
|---|---|
| Mac | /Applications/Tallybook.app/Contents/MacOS/tallybook-desktop --mcp |
| Windows | "C:\Program Files\Tallybook\tallybook-desktop.exe" --mcp (typical install path) |
| Linux | tallybook-desktop --mcp |
It opens your most recently used book, read-only. Add:
--writeto let the agent change the book (categorize, record entries, budgets).--book /path/to/Your.tallybookto open a specific book.
It shares the app’s recent books, license and keychain, so encrypted books open as they do in the app. The examples below use the Mac path; swap in yours.
Claude
Claude Code:
claude mcp add tallybook -- /Applications/Tallybook.app/Contents/MacOS/tallybook-desktop --mcp
Claude Desktop: Settings → Developer → Edit Config, then add to
claude_desktop_config.json:
{
"mcpServers": {
"tallybook": {
"command": "/Applications/Tallybook.app/Contents/MacOS/tallybook-desktop",
"args": ["--mcp"]
}
}
}
Claude Agent SDK (Python):
from claude_agent_sdk import ClaudeAgentOptions
options = ClaudeAgentOptions(
mcp_servers={
"tallybook": {
"command": "/Applications/Tallybook.app/Contents/MacOS/tallybook-desktop",
"args": ["--mcp"],
}
},
allowed_tools=["mcp__tallybook"],
)
OpenAI
Codex (the CLI, the IDE extension, and Codex in the ChatGPT desktop app):
codex mcp add tallybook -- /Applications/Tallybook.app/Contents/MacOS/tallybook-desktop --mcp
or in ~/.codex/config.toml:
[mcp_servers.tallybook]
command = "/Applications/Tallybook.app/Contents/MacOS/tallybook-desktop"
args = ["--mcp"]
OpenAI Agents SDK (Python):
import asyncio
from agents import Agent, Runner
from agents.mcp import MCPServerStdio
async def main():
async with MCPServerStdio(
params={
"command": "/Applications/Tallybook.app/Contents/MacOS/tallybook-desktop",
"args": ["--mcp"],
}
) as tallybook:
agent = Agent(
name="Bookkeeper",
instructions="Use Tallybook to answer questions about my books.",
mcp_servers=[tallybook],
)
result = await Runner.run(agent, "What was my net profit this year?")
print(result.final_output)
asyncio.run(main())
ChatGPT on the web only connects to MCP servers on the internet, and Tallybook deliberately doesn’t put your books on the internet, so it isn’t supported there.
OpenClaw
openclaw mcp add tallybook \
--command /Applications/Tallybook.app/Contents/MacOS/tallybook-desktop \
--arg=--mcp
Then check the connection with openclaw mcp doctor tallybook --probe.
Anything else
Any agent or app that can launch a local (“stdio”) MCP server works the same
way: give it the command and --mcp as its argument. That covers Cursor,
VS Code, Windsurf, Gemini CLI, Goose and most agent frameworks; look for
“MCP servers” in their settings.
Agents that run only in the cloud can’t start a program on your computer, so they can’t use Tallybook directly.
What the agent can do
| Tool | What it does | Needs --write |
|---|---|---|
book_status | The open book, its sides, license, inbox count, recent books | |
open_book | Switch to another book | |
accounts | Chart of accounts with balances (and the ids other tools take) | |
account_register | Every line in one account, with a running balance | |
entries | Journal entries in a date range | |
report | Profit & loss, balance sheet, trial balance, A/R and A/P aging | |
dashboard | The home screen’s numbers | |
review_inbox | Bank transactions to categorize, with suggestions | |
household_month | The month’s budget vs. spending, alerts and wins | |
bills_due | Household bill due dates and whether they’re paid | |
goals | Savings goals and when they’ll be reached | |
records | Contacts, people, invoices, bills, payments, rules, recurring bills, bank connections | |
tax_summary | Schedule C, estimated-tax plan and 1099s (Business + Tax license) | |
categorize | File an inbox transaction, optionally remembering a rule | ✓ |
accept_suggestions | Accept Tallybook’s suggestion for inbox transactions | ✓ |
record_entry | Post a balanced journal entry | ✓ |
void_entry | Reverse an entry (history is kept) | ✓ |
set_budget | Set a line of a month’s household plan | ✓ |
call | Any other API method by name, e.g. invoices.create, bank.sync | for changes |
Money is integer cents (12345 is $123.45) and dates are YYYY-MM-DD.
Guardrails
- Read-only by default. Without
--write, nothing that changes the book is offered, andcallrefuses any method that would. - The book’s own rules still apply. Entries must balance, closed periods stay closed, and voids reverse rather than delete. License gates are the same as in the app.
- Never available to an agent, in any mode: deleting a book, encryption and recovery keys, licenses, cloud backup destinations, and connecting or disconnecting banks. Do those in the app.
- The app and an agent can have the same book open. Each waits briefly for the other’s write to finish. Screens in the app show an agent’s changes the next time they load.
- Your books go to the agent’s AI. Whatever the agent reads is sent to the AI service it uses. Pick one you trust with your finances, and stay read-only unless you want it making changes.