Hermes Agent for accounting earns a 7 out of 10 from me as a bookkeeping prep assistant, and a much lower mark if you hoped it would run the ledger.

The ratings below are my editorial reading of the official Hermes Agent docs and the v0.23.0 catalogs, not lab results.

Each score has a one-line reason you can check against the source.

This is a review of the tool for admin work around the books, and nothing in it is tax, legal or financial advice.

If you'd like my team to sanity-check a set-up for your own firm, book a free SEO strategy session.

The Scorecard At A Glance

CriterionScore /10Why, from the docs
Reading documents8Word and Excel convert out of the box; PDFs need an optional converter that installs on first use.
Spreadsheet prep8The bundled microsoft-office skill edits .xlsx and .csv offline, but not legacy .xls.
Payment data7Stripe, Square and PayPal sit in the Nous-approved MCP catalog; nothing there acts as a bank feed.
Accounting software links2No catalog entry for Xero, QuickBooks, Sage, Wave, FreshBooks or Zoho Books at v0.23.0.
Scheduling9Cron takes plain language, loads skills per job and can pin its own model.
Money-safety controls7Untrusted MCP servers ask before every write call, but YOLO mode skips that prompt.
Outbound email control4No documented approval step before the gateway sends; you rely on wording and skill defaults.
Data control9Self-hosted, one profile per job or client, and local models through Ollama.
Set-up effort5Command line, environment variables and a required sender-authentication pin for email.
Overall7A capable prep assistant with one big gap: the ledger itself.

The pattern is clear: high marks for the jobs around the paperwork, low marks wherever money or the ledger is involved.

That's the right way round for a bookkeeping tool, as long as you know it going in.

I weighted the criteria by what a small firm does most often: collecting documents, preparing sheets and reading payment data come up every week, while set-up happens once.

A firm that cares most about a direct ledger link should read the second-lowest row first and probably stop there.

Everything is scored for Hermes v0.23.0, the stable release from 5 November 2026, so a later release could move a row.

Strength: It Reads The Paperwork You Already Have

The document extraction docs say read_file turns .docx and .xlsx into readable text with nothing extra installed.

PDF, legacy Office and OpenDocument files go through the optional firecrawl-anydoc converter instead.

The honest part I liked: when scanned pages come back empty, the tool says so, with page ranges, rather than handing over a short extraction as if it were complete.

The bundled document-to-action-items skill goes further and requires every extracted field to cite where it came from.

For an accountant checking a client's supplier invoices, that citation trail is the difference between checking and redoing.

Strength: Schedules That Carry Your Own Rules

The cron docs let one job load several skills before it runs, so a month-end job can carry your filing rules and the Office skill together.

Results can go back to a chat, into files or to another platform.

Each job can also pin its own model, so changing the model you chat with doesn't silently change how your month-end runs.

That's why scheduling gets the top score on the card.

Jobs don't have to wait for the clock either: the webhooks docs let an outside event, such as a Stripe payment failing, start a run straight away, with HMAC signature checks.

Scheduled runs are also boxed in by default: dangerous commands are refused in cron, and a job can't create or edit other jobs unless you allow it.

Strength: Payment Data On A Short Leash

The Nous-approved MCP catalog carries vendor-hosted servers for the three payment apps small firms use most: Stripe covers payments, customers and invoices; Square adds its catalog and orders; PayPal adds subscriptions.

The MCP config reference offers two controls that matter for money: an exclude list, shown for Stripe with delete_customer and refund_payment, and trust: untrusted, which makes every write-capable call ask first.

Used together, the agent can read failed payments all day and still can't refund one without a person clicking Allow.

It loses points because those controls are opt-in, and YOLO mode or approvals.mode: off skips the prompt entirely.

๐Ÿ”ฅ Want the scored set-up, not just the scores? Inside the AI Profit Boardroom you get step-by-step video trainings, the installable Agent OS, four coaching calls a week to review your own agent, and 3,400+ members comparing tools like this one. โ†’ Join the AI Profit Boardroom

Strength: You Decide Where The Data Lives

Hermes runs on your own machine or server, and the code is MIT-licensed on GitHub.

The profiles feature gives each agent a separate home with its own settings, keys, memory, skills and scheduled jobs.

The providers page describes local Ollama as a good fit for privacy-sensitive work and offline use.

For a practice holding client statements, that combination is hard for a hosted agent to match on this one criterion.

The docs also warn against running two agents on one profile, which pushes you toward a clean one-client, one-profile layout.

The catch, flagged on the same providers page, is that small local models can run out of context once the full toolset and system prompt load, so local set-ups need a trimmed toolset.

Limit: There Is No Ledger Connector

This is the score that drags the total down.

Neither the plugin catalog nor the MCP catalog at v0.23.0 has an entry for the accounting packages small firms usually run.

The Plaid entry might look like a bank connection, but its manifest describes the Plaid developer dashboard for debugging integrations.

In practice, you export from the bank and the accounting app, the agent prepares, and a person imports the result.

Hermes can connect to any MCP server you add yourself, so a firm with a developer could wire one up, but that is your own build, not something Nous has reviewed.

For this scorecard, only what the catalogs list counts, and for the ledger that is nothing.

Limit: Outgoing Email Is Not Gated

The security page describes approvals for dangerous shell commands, and the MCP reference adds them for writes on untrusted servers.

I found no docs line describing an approval step before the email gateway sends a reply.

The bundled email-inbox-triage skill defaults to drafting rather than sending, which helps.

Even so, the safe practice is to write "draft only" into every job that touches a client and let a person send.

For a tool that will be chasing clients about money, that earns a 4.

Limit: Scanned Paper And Set-Up Effort

Paper receipts scanned as images are the everyday weak point, since PDF conversion reads text layers only.

The docs' fixes are rendering pages for the vision tool or a bulk OCR skill with a large install.

Set-up asks for more than most owners expect: environment variables for the mailbox, an app password, and an authentication pin without which the gateway drops mail.

A firm with no technical help will stall here, which is why set-up effort scores a 5.

Real-World Fit: A Bookkeeper's Month-End Prep

  1. Clients forward documents to the agent's dedicated mailbox, which accepts only listed senders.
  2. A daily job names and files each document and adds a row to the prep workbook.
  3. On the 1st, a second job builds a pack of unmatched items and missing documents for each client profile.
  4. The bookkeeper reviews the pack, fixes what's wrong and does the posting in the accounting software.

Every step there plays to a high-scoring row on the card, and none of it needs the missing ledger connector.

The weak rows only bite if the bookkeeper lets the agent email clients directly, which the set-up should rule out from day one.

One practical tip from the email docs: the gateway ignores noreply-style senders, so clients have to forward documents themselves rather than auto-forwarding from a billing system.

Real-World Fit: A Cafรฉ On Square

  1. The owner connects Square from the MCP catalog with writes set to ask first.
  2. A Monday job summarises last week's orders and payments by day.
  3. The owner checks the summary against the till report and notes any gaps.
  4. Supplier receipts photographed on a phone go to the agent's mailbox and are read by the vision tool.

Good fit, with one condition: the person who reads the summary still owns the numbers.

The cafรฉ never needs the ledger connector or outgoing email at all, which is why it lands among the best matches for the tool as it stands.

๐Ÿ“ž Want a straight answer on whether this fits your business? Book a free SEO strategy session and my team will look at your workflows and tell you what to automate first. โ†’ Book your free call with my team

Who Should Use It

  • Bookkeepers and small practices that want document intake, prep workbooks and chase-up drafts on a schedule.
  • Shops and cafรฉs on Stripe, Square or PayPal that want a weekly read-only summary of payments.
  • Owners who care where client files are stored and want the option of a local model.
  • Anyone with a technical person who can spend an afternoon on set-up.
  • Practices that want each client kept in a separate profile with its own rules and schedule.

Who Shouldn't

  • Firms that want an agent to post, reconcile or close inside their accounting software.
  • Anyone hoping the agent will pay suppliers, issue refunds or file returns.
  • Teams with nobody willing to edit a config file.
  • Owners who'd rather have a hosted agent with draft-before-send email built in; our Grok Bot for accounting verdict scores that route.

Final Verdict

Hermes is a 7: excellent at reading, sorting, scheduling and summarising, deliberately weak at touching money, and missing the ledger link entirely.

Set it up read-first, keep sends in human hands, and it does the dull half of bookkeeping well.

Ask for the other half and you'll be disappointed, because the docs never promise it.

The AI Profit Boardroom covers the build on video if you want to follow along, and my team can review your plan on a free call.

Also On Our Network

FAQ: Hermes Agent For Accounting Verdict

What overall score does Hermes get for accounting?

A 7 out of 10 as a prep assistant, based on an editorial reading of the docs. It scores high on documents and scheduling and low on ledger links and outgoing email control.

What is the biggest weakness?

There is no catalog entry for the accounting packages small firms use, so exports and imports between the agent and your ledger stay manual.

Is the Plaid entry a bank feed?

No. Its manifest describes the Plaid dashboard for integrations, Items and usage debugging, not bank transaction syncing.

Were these scores from benchmarks?

No. They are editorial ratings from the official docs and catalogs, each with a reason you can check in the source.

๐Ÿ“บ Video notes + links to the tools ๐Ÿ‘‰

๐ŸŽฅ Learn how I make these videos ๐Ÿ‘‰

๐Ÿ†“ Get a FREE AI Course + Community + 1,000 AI Agents ๐Ÿ‘‰

About Julian

I'm Julian Goldie, an SEO entrepreneur, author and founder of the AI Profit Boardroom, which has 3,400+ members.

I help business owners scale with AI agents, automation and SEO.

  • I've built a 7-figure SEO and link building agency, Goldie Agency, with a team of 70+ SEO experts.
  • I've grown a YouTube channel to 400,000+ subscribers.
  • I wrote "Link Building Mastery".
  • My Udemy courses have taught over 29,000 students.

โ†’ Get my best AI training inside the AI Profit Boardroom

Score your own paperwork against this card before you install anything, and book a free strategy session if you want a second opinion on Hermes Agent for accounting.