Installation / prerequisites
- Beancount.io
- Supported, observed: install bea 0.1.0; no account for local commands
- hledger
- Supported: single binary via package manager or Docker (1.52.3)
The 2026 agent accounting guide
Find the right accounting tool for your AI agent.
Five tools. Seven workflow criteria. Start with where your books live and what your agent needs to do.
Five tools. Five different fits.
This order prioritizes agents that draft ledger changes: observed preview and repeat-safe imports, then documented workflows, then cloud access and read-only-by-default tools. Choose the fit for your books.
- Observed run
Beancount.io
Free core · local files, no account needed
Local plain-text double-entry books with a CLI built for agents: preview before apply, repeat-safe imports, and the only observed end-to-end run on this page.
Best for:Agent-builders who want every ledger change previewed as text and verified before it lands.
Strengths
- Observed preview-first run: 3 ready, 0 written, ledger bytes preserved
- Repeat apply writes 0 with 3 exact duplicates; an unbalanced write is rejected
- No account, local files, and machine-readable JSON output
Trade-offs
- You own the setup: chart of accounts and hosting are yours to arrange
- Your accountant may need exported reports rather than direct access
- Documentation reviewed
hledger
Free, open source
Local plain-text accounting with a documented dry-run and idempotent import workflow, maintained since 2007.
Best for:Plain-text scripting veterans who want local double-entry without a new platform.
Strengths
- Documented import with a dry run plus idempotent re-import
- Local journal files — no accounts, no cloud
- Reports to terminal, HTML, CSV, JSON or SQL
Trade-offs
- The agent path is the documented CLI, not an observed agent run
- Plain-text setup suits technical users most
- Documentation reviewed
Xero
Cloud ledger · developer account and OAuth required
A hosted Remote MCP server, generally available, plus an open-source local server, CLI and agentic SDK — read and write tools scoped by OAuth grants against your Xero org.
Best for:Agents operating inside an existing Xero organisation.
Strengths
- Official read and write tools: invoices, contacts, payments, journals and more
- LLM-oriented TOON output alongside JSON and CSV
- Hosted Remote MCP: nothing to install, OAuth with PKCE
Trade-offs
- Needs a Xero account, a developer app and granted scopes; Remote MCP on a live organisation needs the Core plan or above
- Remote MCP V1 writes ask for no confirmation; Xero advises building your own step
- Documentation reviewed
QuickBooks Online
Cloud ledger · Intuit sign-in or developer app
Intuit's hosted QuickBooks connector for Claude (US) and ChatGPT, plus an official open-source MCP server with full entity CRUD and financial reports.
Best for:Agents operating inside an existing QuickBooks Online company.
Strengths
- Full create, read, update, delete and search over 29 entity types plus 11 reports
- Flags to run read-only by disabling write, update or delete categories
- Hosted Claude connector: Intuit sign-in, no developer app, destructive actions confirmed
Trade-offs
- The hosted connector is US-only; the local server needs an Intuit developer app and a browser OAuth handshake
- A general preview-before-write is unverified in the sources checked
- Documentation reviewed
Puzzle
Cloud ledger · MCP connector, read-only by default
Read-only MCP access to a Puzzle company by default, with confirmed writes as an opt-in Beta — a Q&A window first, not a bookkeeping backend.
Best for:Conversational questions over a Puzzle company in Claude, ChatGPT, Cursor or Codex.
Strengths
- Lightest setup of the cloud options: a connector URL, no install
- Works from Claude.ai on any plan, ChatGPT, Cursor and Codex
- When writes are on, each change is confirmed with you before it is made
Trade-offs
- Read-only by default; writes are a Beta you turn on per connection
- Rate limited, with no dry run or repeat-import behavior documented
The same seven criteria, all five tools
Every tool is assessed against the same criteria. “Unverified” means the sources did not cover it; it does not mean the feature is missing.
Authentication
- Beancount.io
- None needed locally; hosted credentials only for bea ask
- hledger
- None needed: local files, no accounts
Read / write scope
- Beancount.io
- Reads and writes, observed: query, report, preview-then-apply import
- hledger
- Reads via reports; writes by appending and via import
Local vs cloud data
- Beancount.io
- Local: plain-text ledger files
- hledger
- Local: plain-text journal
Preview / validation
- Beancount.io
- Observed: preview 3 ready / 0 written; check valid; unbalanced write rejected
- hledger
- Documented: import dry run; entries balanced and validated; check command
Structured output
- Beancount.io
- Observed: JSON envelopes with rows and columns, exit codes, BQL
- hledger
- Reports to terminal, HTML, CSV, JSON or SQL
Repeated imports
- Beancount.io
- Observed repeat-safe: reimport writes 0, bytes preserved
- hledger
- Documented idempotent re-import via hidden state files
| What matters | Beancount.io | hledger | Xero | QuickBooks Online | Puzzle |
|---|---|---|---|---|---|
| Installation / prerequisites | Supported, observed: install bea 0.1.0; no account for local commands | Supported: single binary via package manager or Docker (1.52.3) | Hosted Remote MCP: nothing to install, Core-plan developer app; local server and CLI: Node 18+ and API credentials | Hosted connector: no install, US QuickBooks Online subscription; local server: clone, build, Intuit developer app | No install: Puzzle account, MCP client and connector URL |
| Authentication | None needed locally; hosted credentials only for bea ask | None needed: local files, no accounts | Xero OAuth 2.0: PKCE for Remote MCP and CLI, or a Custom Connection for the local server | Hosted: Intuit sign-in and permission grant; local: OAuth2 app handshake, production needs public HTTPS | Puzzle OAuth for chosen companies; stays signed in if used once every 30 days |
| Read / write scope | Reads and writes, observed: query, report, preview-then-apply import | Reads via reports; writes by appending and via import | Remote MCP: 18 read tools and 21 create, update, authorise and void tools, scoped by OAuth grants | Hosted: reports, invoices, estimates, customers, transaction imports; local: full CRUD and search, writes can be disabled | Read-only by default; write Beta per connection: categorize, bills, invoices, journal entries, reconcile |
| Local vs cloud data | Local: plain-text ledger files | Local: plain-text journal | Cloud: hosted by Xero or run locally; ledger lives in the Xero org | Cloud: hosted connector or local stdio subprocess; ledger lives in the QBO company | Cloud: hosted MCP endpoint over Puzzle company data |
| Preview / validation | Observed: preview 3 ready / 0 written; check valid; unbalanced write rejected | Documented: import dry run; entries balanced and validated; check command | Remote MCP V1 asks no confirmation (Xero advises adding one); CLI validates file input; dry run unverified | Hosted: destructive actions need confirmation; local: schema validation; dry run unverified | With writes on, each change is confirmed before it is made; dry run unverified |
| Structured output | Observed: JSON envelopes with rows and columns, exit codes, BQL | Reports to terminal, HTML, CSV, JSON or SQL | JSON, CSV and TOON output modes | MCP tool results plus invoice PDF download | Unverified: conversational answers in the MCP client; result envelope not covered |
| Repeated imports | Observed repeat-safe: reimport writes 0, bytes preserved | Documented idempotent re-import via hidden state files | Unverified in the sources checked | Unverified in the sources checked | Unverified: no import path documented |
Frequently Asked Questions
Which accounting tool is best for an AI agent that drafts ledger changes?
Beancount.io is the only tool on this page with an observed preview-first agent run behind it: the agent previews 3 ready rows, applies 3 entries, and a repeat apply writes 0 while the ledger bytes are preserved. hledger documents the same shape of workflow — dry-run import plus idempotent re-import — from its official manual. Xero and QuickBooks Online give agents official read and write tools behind their cloud OAuth setups, which fits agents operating inside an existing cloud org.
Which accounting tool lets an AI agent re-run the same import safely?
Beancount.io and hledger. On Beancount.io the behavior was observed: an identical repeat apply reports 0 ready with 3 exact duplicates, writes 0 entries and leaves the ledger bytes unchanged. hledger documents idempotent import in its manual, keeping state in a hidden .latest file so a re-run appends only new transactions. For Xero, QuickBooks Online and Puzzle, what a repeated import does is unverified in the sources checked. Agents retry, so test this before you let one import unattended.
Does an AI agent need an account to work with a local ledger?
No. Beancount.io's local import, query, report and check commands run on plain-text files with no account; only bea ask needs hosted credentials. hledger needs no account at all. Xero, QuickBooks Online and Puzzle keep the ledger in the cloud, so an agent first needs that service's account and an OAuth grant, whether it connects through a hosted MCP server or a local tool.
Who publishes this comparison?
Beancount.io publishes this page and is one of the five tools compared. Beancount.io cells link to an observed, reproducible run; competitor cells cite official documentation checked on 2026-09-10, with the Xero, QuickBooks Online and Puzzle details rechecked on 2026-11-01. No competitor was runtime-tested here, no paid account was used, and unknowns are labeled unverified rather than treated as failings.
See what an agent can do with a ledger.
Install bea, hand your agent one task, and check the result yourself. Local books need no account.