Plain-Text Books vs Books by Mercury — Research Dossier

Asked by Devon on 12 October 2026, for the question of how the businesses keep their books from here. Six questions: where the plain-text accounting tools stand now and whether hledger is still the best fit; what Books by Mercury and Ramp’s accounting actually offer; what the homelab already has from last year’s catch-up; what a weekly agent catch-up costs against Mercury; how each option fits owning the data; and which option to take. Sources at the end. Vendor statements are marked as claims. No amounts or ledger content appear here.

1. The plain-text landscape, October 2026

GitHub activity for the 12 months to 12 October 2026 (queried 12 October 2026 through the GitHub API). The hledger repository now lives at hledgerorg/hledger.

What the numbers say:

• hledger is the most active plain-text accounting project, shipped 17 release tags in the year, and its latest stable release is 1.52.4 (10 September 2026). Its weakness is the same as the others’: one maintainer does most of the work.
• Beancount v3 has a slow core. Since the 2024 split, importers, queries and prices live in separate projects, which is the extra work the earlier Main Branch comparison counted against it. Fava is healthy.
• rustledger is a Beancount-compatible Rust rewrite with an MCP server and an agent-oriented command line. It is nine months old with one maintainer.
• hledger has no official MCP server. Several community ones exist and the most starred has 65 stars (checked 12 October 2026). Its stable machine interface is -O json.

Researcher’s judgment. For CSV rules built in, one repository per business, and agents driving a command-line tool, hledger is still the strongest choice. Nothing found this year beats it. rustledger is worth watching for its agent ergonomics but is too young and too dependent on one person to adopt.

2. Books by Mercury and Ramp

Everything in this section is a vendor claim unless marked otherwise.

Books by Mercury, from its launch post (16 September 2026) and product page, both read 12 October 2026:

• Claim: price is $35 a month, and free for all of 2026. Devon had $30 in mind; the pages say $35.
• Claim: no transaction or entity limit is listed. The only cap mentioned is that advisor seats are uncapped. Nothing says the price is capped. Per-entity pricing is not stated, and a Mercury account is required, so the price may well apply once per business (unconfirmed).
• Claim: a full double-entry engine, cash or accrual, automatic categorisation and reconciliation, profit and loss, balance sheet and cash flow, connections to outside banks, Stripe, PayPal and Gusto, and unlimited bookkeeper seats. It says it is built to replace QuickBooks, not sync with it.
• Claim: a full accounting API and an MCP layer for agents are coming “soon”. Neither is shipped.
• Not found on those pages: receipts, multi-entity consolidation, tax handoff, export formats.
• Export, from Mercury’s support site (searched 12 October 2026): transactions export as CSV or XLSX, categorised transactions with their category codes export as CSV, statements download as QuickBooks or NetSuite CSV, and manual journal entries export as XLSX only. A full general-ledger or journal export was not found on these pages. Mercury’s own move-in guide says Books does not import history from QuickBooks or Xero.

Ramp, from ramp.com/pricing (read 12 October 2026), all claims: a free tier with basic accounting rules and QuickBooks and Xero sync, Plus at $15 per user per month plus a team-size platform fee (multi-entity, AI field coding, NetSuite and Sage Intacct, scheduled exports), and custom Enterprise. No standalone accounting agent is described. Ramp syncs to a ledger; it is not one. Its API documentation was not read.

Researcher’s judgment. Mercury’s product is credible and the free year costs nothing to try, but a ledger leaving intact is not yet something it documents, and the two things that would make it agent-friendly are promises.

3. What we already have

• Main Branch chose hledger on 11 May 2026, after a written comparison with Beancount v3, Ledger and CSV or SQLite. The wrapper commands for checking and status ship now; the full books command line is Main Branch issue 128.
• Each business, Awake Happy and Noontide among them, has its own hledger repository.
• The operating contract is the books workspace coldstart file. Its core rules: bank statements and provider exports are the only evidence, classification lives in rules files and not scripts, the repository check must pass before any push, and raw evidence sits in a checksummed archive on the home server.
• That doctrine exists because bookkeeping apps editorialised: the errors found in 2026 each came from an app and each catch came from a statement.
• A fresh-agent adversarial audit protocol checks every deliverable for the accountant against the repository and the statements before it is sent.
• Timing: the commit history shows bursts of a few intense days per business inside a two-month window. That fits Devon’s recollection of under a week per business; it does not prove one week for all.
• A Mercury agent-tools plan was not found in the homelab docs, fleet notes, personal state or the books repositories.

4. Cost model

All figures are estimates, ranges and not measured.

• Mercury: free through 2026, then $35 a month, or $420 a year, per business if it is charged per business. Businesses banking elsewhere would need the outside-account path.
• Weekly agent catch-up: with mature rules most rows categorise by themselves and the leftover is small, so a few hundred thousand tokens per business per week, which is a few dollars at most across the businesses per week. Devon’s guess of about $5 a week sits inside this. This is the researcher’s judgment from the shape of the work; no session usage was summed and no price list was consulted.
• Maintenance hours: roughly one to three hours a month in total for tool upgrades, rules drift and checks.
• Real cost is attention: statements still have to be downloaded, and exceptions reviewed.

Researcher’s judgment. The agent route costs less cash than Mercury once there is more than one business. Mercury costs less attention if its categorisation is trusted, which the evidence doctrine says not to do by default.

5. Fit with the principles

A hybrid is possible: Mercury as the live tool, with a weekly statement-anchored hledger export in git. It becomes real only when Mercury documents a journal or API export. Until then Mercury is the bank and a download source, and hledger stays the book.

6. Recommendation

Keep hledger as the book of record. This confirms the Main Branch decision of 11 May 2026; it does not supersede it. Use the free Mercury Books year as an observation tool, never as truth.

First three steps:

1. Finish Main Branch issue 128 (the full books command line) and add a one-prompt weekly catch-up per business, so the weekly agent run is one command.
2. Run Mercury Books read-only on one business for a month and compare its categories against the statement-anchored hledger book.
3. Revisit by 1 December 2026: has Mercury shipped a documented journal or API export, and what is the per-business price?

What would change the answer: a documented ledger export plus API from Mercury, the accountant preferring its reports, the catch-up needing more than about three hours a month of attention, or the hledger maintainer risk materialising (fallback: hledger writes Beancount, and rustledger is the watched alternative).

Sources

• Introducing Mercury Books (16 September 2026): price, free period, engine, coming API and MCP.
• Mercury Books: price, advisor seats, accrual or cash, external connections.
• Mercury support: importing transactions with a CSV: CSV paths.
• Mercury support: creating and managing journal entries: journal entries export as XLSX.
• Mercury support: moving to Books from QuickBooks Online or Xero: no history import.
• Ramp pricing: plans and accounting features.
• hledger on GitHub: activity, releases, issues.
• Beancount on GitHub: activity and issues.
• Fava on GitHub: activity and issues.
• beangulp on GitHub: activity and issues.
• Ledger on GitHub: activity, releases.
• rustledger on GitHub: newcomer, MCP server, maintainers.
• hledger MCP search results on GitHub: the most starred community MCP server.
• Main Branch report “hledger vs Beancount fit for mb books” (11 May 2026): the earlier decision and its scorecard.
• Books workspace operating and audit files (read 12 October 2026): evidence doctrine and audit protocol.