hledger

What it is

hledger is plain-text accounting software in Haskell, created by Simon Michael and maintained since 2007, with 140+ contributors. It reads a journal you write (largely compatible with Ledger), checks it and produces reports. It ships three interfaces: the CLI, hledger-ui in the terminal and hledger-web in the browser, which also serves a JSON API. It is free software under GPL-3.0-or-later.

It is not a bank-sync service, has no hosted version and keeps no history of the journal beyond what you put in Git.

How it works

You record transactions in a journal, or import them. hledger import reads CSV, TSV or SSV through a .rules file that maps columns to dates, amounts, payees and accounts, with conditional rules for categorizing. Since 1.50 (2025-09-03) a source rule can run a command to produce the data (the docs show fetching JSON from SimpleFIN and converting it to CSV), and an archive rule saves a dated copy of every imported file in data/ next to the journal, “useful for troubleshooting, detecting variations in your banks’ CSV data, regenerating entries with improved rules”. import --get runs your fetch scripts first; --dry-run previews.

The weak point is how import decides what is new. The manual says it “remembers the latest date processed previously in each CSV file” in a hidden .latest.FILE and skips records before that date (and the first N on that date). It works when the file name, item dates and same-date order are stable and the newest items have the newest dates. It deliberately does not deduplicate identical records, since “it’s generally not possible to reliably detect duplicates in bank CSV”. A bank that backdates a pending item, reorders a day or renames its export breaks the assumption without an error.

Data and access

You hold the journal. Reports export as txt, html, csv, tsv, json, sql, ledger or beancount. The hledger-web API offers GET routes for transactions, accounts, prices and commodities and a PUT /add that takes a complete transaction. Access is set with --allow (view, add or edit; default add), and the manual is direct: “There is no access control.” You bind it to localhost or put an authenticating proxy in front.

Agents

There is no official MCP server or AI feature. The JSON output and API are a reasonable base for an agent, and plaintextaccounting.org lists accountant24, a local-first AI agent for hledger, and a guide to using Claude Code with hledger in Obsidian. Writes have no approval gate: an agent can append to the journal or call /add, and hledger check --strict validates afterwards.

The development side is notable. hledger 1 (up to 1.52.1) was written without AI assistance; hledger 2 (from 1.99.1, 2026-03-28) is “exploring careful AI-assisted development”, with every change “steered, reviewed, and tested by Simon or a contributor”, a ban on AI tools for first-time contributors’ PRs and published token estimates. The project used an Anthropic open-source credit in 2026 Q2 and Q3.

Customers

Individuals, freelancers and some small businesses, most of them comfortable with a terminal. About 4.8k GitHub stars; plaintextaccounting.org (2026) counts about 199 committers. Open Collective shows $12,526.08 raised in total and an estimated $937 annual budget (2026-10-10). Users praise its speed, documentation, error messages and CSV rules; the costs are the learning curve and rules per bank.

History

0.1 was announced 2008-10-15 as “a command-line accounting tool similar to John Wiegley’s c++ ledger”. 1.0 came 2016-10-26. The 1.5x line added import archiving and data commands. The 2.0 previews (1.99.x, latest 1.99.5 on 2026-10-01) add lot tracking, capital gains, a holdings report and commodity aliases, with stable 1.52.4 (2026-09-10) alongside.

For ownpurse

hledger is kin and a likely neighbour for the same people: someone who already keeps an hledger journal could point it at ownpurse output. Worth borrowing: declarative rules per source, keeping the original file on every import, a dry run before any write, and many honest output formats.

The lesson to avoid is the identity rule. A date high-water mark is cheap and usually right, and wrong in exactly the cases that matter: backdated items, reordered days, identical purchases. ownpurse files each download by hash and identifies records by source id, natural key or fingerprint, so reloading changes nothing. And where hledger-web lets anyone who can reach it add transactions, ownpurse keeps reads open and writes behind a sealed package a person approves.