Meridian Starter Review (2026): Your AI Agent, Your QuickBooks and Xero Books, and No One at the Door

Pilot launched Meridian Starter on 20 October 2026: one workspace where Claude, Claude Cowork, ChatGPT or Codex reads from — and writes to — every client ledger you connect. Free for one person, $50 per extra user or company, and no approval step documented before the agent changes the books.

By Bookkeeping Reviews · Published October 22, 2026 · Last updated October 22, 2026 · 17 min read

What Is Meridian Starter?

Meridian Starter is a shared workspace for doing client accounting work with an AI agent that is connected to the ledger. It launched on 20 October 2026, announced from San Francisco by Meridian by Pilot, and it is available now at starter.meridian.pilot.com. The distinction matters, because the market is full of connectors: Starter is not selling the cable, it is selling the desk around it — the client list, the checklist, the context notes, the teammates, and the decision about how much of each ledger the agent may touch.

The company behind it is worth understanding, because it explains the product's instincts. Pilot is an accounting firm, not a software vendor that added accounting later: it has run bookkeeping, controller and CFO work for startups and small businesses since 2017, and that operational history is the asset. Meridian is Pilot's AI platform for accounting firms — the release describes it as built inside Pilot and refined across 8,000 businesses and more than 200,000 months of books — and it automates the month-end close across a whole portfolio, with AI categorization, reconciliation and tick-and-tie, for firms running 50 to 5,000 clients. Meridian Starter is the entrance to that: one client, one task, one agent, at a price a solo bookkeeper can start on today.

The gap it is aimed at is stated in the launch release, and it is a fair description of how most accountants actually use AI in late 2026:

"AI use among accountants is now mainstream, with 98% reporting they use it to support their clients, most often with Claude or ChatGPT. That work usually happens one chat and one company file at a time, and what the AI learns about a client stays in one person's chat history."

That is the whole thesis, and it is correct. A bookkeeper who has spent a year teaching Claude how a particular client codes its software costs has that knowledge in one thread, on one machine, in one person's head. Starter's answer is to move that knowledge into a place the firm owns: company context per client, checklists built from templates and instantiated per client, and a team with roles that all work from the same books and the same notes.

What is actually in the workspace

• Ledger connections. QuickBooks Online and Xero companies in a single list, each with its own access level. The site's own product shot shows one QuickBooks company marked Read & write and one Xero company marked Read only, side by side — which is the clearest statement of the design philosophy you will find on the page.
• Checklists. Write the steps of a repeatable job once, save it as a template, then start it for each client and have the agent work through it against that company's books and context.
• Shared context. Accounting policies, preferences and background per company, available to you, to anyone you invite, and to the agent — so the instructions stop being re-explained in every chat.
• Team access. Invite teammates and assign roles so several people work from the same connections instead of each maintaining a private set.
• Agents. Claude (including Claude Cowork), ChatGPT and Codex, connected through separate QuickBooks and Xero agent connections. Meridian's reason for keeping them separate is sensible: each ledger has its own concepts and capabilities, and a narrower toolset is easier for a model to use correctly.

Onboarding is four steps and takes about as long as it sounds: create a free account (sign in with Intuit, Xero or Google), connect companies and choose each one's access level, add the QuickBooks or Xero connection to Codex or Claude Cowork, then give the agent work.

What Your Agent Can Actually Do in the Books

Meridian's answer to "what can it do" is one sentence long and it is the most important sentence on the site: "Basically anything that can be done via the QuickBooks and Xero API can be done by your agents using this connector." That is a much larger claim than the list of examples that follows it, and buyers should treat the claim, not the examples, as the specification.

What it reads

• Transaction-level detail. This is the differentiator Meridian names explicitly, and it is the one that matters: it says the official integrations "generally can't read transaction-level details (it can only run reports)". An agent that cannot see individual transactions can only summarize; an agent that can see them can work.
• Statements compared against the ledger. The site's first example prompt is "Compare this bank statement with Xero and list anything missing" — reconciliation support rather than reconciliation automation.
• Accounts, contacts and reports for the connected company, which is what makes "categorize last month's expenses and ask me about anything unusual" possible.
• Your own context and checklists, which live in the workspace rather than the ledger, and which the agent reads alongside the books.

What it writes

• Categorize expenses — the highest-volume, lowest-judgement job in bookkeeping, and the one where an agent's error is cheapest to fix if you catch it in the same week.
• Create invoices — including from a spreadsheet, one of the four example prompts on the page.
• Create journals — journal entries, which Meridian also names as something the official integrations cannot do.
• "Arbitrary other objects." Meridian's own words in the FAQ, in the same breath as journal entries. Whatever the API exposes, the agent may touch.
• Void and delete, implicitly. Neither the site nor the release excludes them, and both are things the QuickBooks and Xero APIs can do. Compare the Financial Cents MCP, which states that deletes are not available over the connection at all. Meridian has not made that promise, so do not assume it.

The other thing to understand is what Starter deliberately does not do. It does not automate a close: the release reserves recurring production — categorization, reconciliation and tick-and-tie across a portfolio, with deterministic accounting checks — for the full Meridian platform. Starter is one task at a time, directed by a person, which is the point of an entry product but also the reason its value depends entirely on how well you brief the agent.

Pricing: Free for One, Then a Meter That Counts Twice

Meridian's own FAQ answers the price question in a way that is unusually clear for this category, so here it is almost verbatim, followed by what it means in monthly money.

The second thing to understand is that nothing in that price includes the AI. Meridian connects your Claude, ChatGPT or Codex account; it does not sell you one. Custom connectors are a paid-tier feature on the major assistants rather than a free one, and Claude Pro and ChatGPT Plus have sat around $20 per person per month at the tiers we last checked. So a three-person firm is not paying $1,050 a month for Meridian Starter; it is paying $1,050 plus three assistant subscriptions, plus QuickBooks or Xero as before. The full stack is what you should compare against, not the Meridian line.

Worked examples, with the arithmetic shown

Read the third row carefully, because it is where most practices will land and where the pricing stops looking like a free product. $1,050 a month for twenty clients is $52.50 per client per month — roughly the price of a dedicated per-client bookkeeping tool, before you pay for Claude, and before you pay for the assistant seats. At five people and fifty clients the effective rate is about $53 per client per month, which is the plateau the model settles on: the per-company fee dominates everything once the client list grows past a handful.

The Approval Question

This is the section that decides whether Meridian Starter is right for you, and it is short because the evidence is short: Meridian documents no confirmation step before the agent writes to your books.

The site carries around a dozen FAQ entries — who it is for, whether it is free, team sharing, whether it replaces QuickBooks, what the agent can do, read-only access, supported agents, why the connections are separate, security, data retention, why not the official integrations, feature requests — and not one of them describes the agent stopping to ask before it changes something. The only control described is the access level: read-only or read-write, chosen per company, when you connect it. That is an authorization, not an approval. It answers "may this agent touch this client at all" and says nothing about "may this particular write happen".

Why this stands out, this month of all months

Two products shipped within days of Starter and both chose the other default:

• Xero's connected app for Google Gemini states that any operation which modifies financial data — issuing an invoice, sending a payment reminder — requires explicit user confirmation in the connected app before it executes, with the review step and final judgement left to the accountant or bookkeeper. That is the approval gate as a product principle. Our Xero × Gemini review sets out the whole design.
• The Financial Cents MCP, launched 20 October 2026, says the assistant confirms before it makes any change, that deletes are intentionally not available over the MCP at all, and that every tool call is audited and rate limited.

Meridian has taken a third position, and it is not an accident — it follows from what the product is for. An agent that must ask permission for every line cannot categorize forty invoices unattended; the value is precisely that it finishes the job and reports back. Meridian's own framing is that agents are most powerful when they can do anything a person can, and a confirmation dialog on every write is the opposite of that. If you accept the premise, the missing gate is the price of the capability.

The gate that may exist anyway — and why you should not count on it

One honest qualification, because we have not connected a ledger and cannot tell you what the screen looks like. Meridian's connection runs inside an AI client — Claude Cowork, Codex, ChatGPT — and those clients have their own tool-permission prompts. It is entirely possible that each Meridian tool call surfaces an "Allow" prompt in the assistant, in which case a gate exists, just not one Meridian owns or describes. Two problems with relying on that:

1. It belongs to the assistant, not the bookkeeping tool. Its behaviour can change with a client update, and it is configured per person, so the gate is only as consistent as your least careful teammate's settings.
2. It is designed to be turned off. Every one of these clients offers "always allow" for a tool, because that is what makes agent work usable. The first time an approval prompt interrupts a forty-invoice batch, someone will click it, and the permission becomes permanent.

How to use it without getting hurt

• Connect everything read-only first. Run a month of questions against read-only ledgers. If the answers are good, you have learned something without risk.
• Grant read-write one client at a time, to the client whose books you know best, and only for the period you are actively working.
• Use a test company. QuickBooks and Xero both let you run a sandbox or a duplicate file. Break something there and learn what breaking looks like.
• Review the ledger's own audit trail daily, not the agent's summary of it. QuickBooks' audit log and Xero's history are the record of what actually changed; the chat is the agent's account of what it thinks it did.
• Small batches, one checklist at a time. Twenty entries you can check by eye; four hundred you cannot, and that is when a systematic error hides.

Security Claims, Read Critically

Meridian's security answer is four short paragraphs in the FAQ. They are better written than most — specific, falsifiable, and free of the word "bank-grade" — but they are the vendor's own words with no audit attached, so here they are with what each one does and does not cover.

The structural risk nobody puts in a comparison table

Meridian Starter is a third-party connector. It reaches QuickBooks Online and Xero through those platforms' APIs, under those platforms' developer terms, and it is not an official Intuit or Xero product — Meridian says as much when it explains why it built its own connections. That has three consequences you should price in:

• It can break without warning. API deprecations, scope changes, rate limits and app-review policies are all set by Intuit and Xero. A connector is a bet that the platform keeps the door open.
• It is one company. The workspace, the connectors and the token store are all Pilot's. If the product is repriced, repositioned or retired, your checklists and client context go with it.
• Nothing about it is a partnership announcement. The release names no collaboration with Intuit or Xero, and the FAQ's defence of the official integrations is Meridian's opinion of them, not a fact either platform has endorsed.

None of that is disqualifying — every serious connector in this category carries the same exposure, including the ones we recommend. It is the reason to keep your context and checklists exportable, and to keep your actual bookkeeping knowledge in your own documentation rather than only in a vendor's notes field.

Pros & Cons

5 Best Meridian Starter Alternatives

Comparison Table

Which One Should You Choose?

If you are one person keeping your own or a few clients' books, start with Meridian Starter today. It costs nothing, it will read transaction-level detail that the official connections will not, and the checklist-and-context structure is the part you will keep using even if you later change agents. Connect one company read-only, ask it a hard question, and judge the answers. That is a genuinely free experiment and there is no reason not to run it.

If you are a two-to-five person practice, do the arithmetic before you invite the second person. The workspace is better than anything else in this category for coordination, but the meter is per person and per ledger, so twenty clients puts you at $1,050 a month plus assistant seats. At that level, compare against the full Meridian platform, which automates the close rather than assisting it, and against doing the coordination in your practice-management tool and using Meridian's free connectors directly. Starter earns its fee when several people genuinely need the same context and the same checklists for the same clients — not merely because it is pleasant to use.

If you are a larger firm — fifty clients and up — Starter is not the product you want. It is deliberately one task at a time, directed by a person, with none of the deterministic accounting checks or portfolio automation. The economics stop making sense around the same point where the capability does. Ask Pilot about the full Meridian platform and price that instead; Starter's value at that size is as a pilot for two or three clients, to see whether the agent's work is good enough for your standards.

Do not choose Starter if any of these is true:

• Your engagement letters or professional-indemnity cover require a human confirmation before any posting. Nothing documented in Starter meets that test. Use Xero × Gemini, or keep Meridian read-only.
• You are Xero-only and want the work inside the platform you already trust. Gemini's connected app is less capable and far more conservative, which for client books is usually the right trade.
• You want the books closed for you. That is the paid Meridian platform, or a managed service like Zeni, not Starter.
• You are outside the US, UK, Canada or Australia, or you work in a non-English ledger. QuickBooks Online and Xero cover much of the world, but the product, the support and the agent workflows here are English-first and built around those two platforms' API surfaces.
• You cannot yet write a checklist. If your month-end is not documented, the agent has nothing to follow. Document the process first; the checklist field will still be there next month.

Data and Responsibility: the Rules to Set Before You Connect Anything

Everything in this section is our advice rather than the vendor's, and it applies to every agent with ledger access, Meridian included. Write these down before the first connection, because after the first write they are retrofits.

1. Snapshot before you grant write access. Export a trial balance and a transaction list for the period the agent will touch, and keep the file. It takes two minutes and it is the only true "undo" you will have, because an agent's mistakes are not one transaction but a pattern.
2. Keep the ledger's audit trail as the source of truth. QuickBooks' audit log and Xero's history record who changed what and when. Read that, not the agent's summary. If a change cannot be traced to a person or an agent session, you have a process problem.
3. Close periods you have filed, and lock them. Both platforms let you restrict edits before a date. An agent with write access and no period lock can reopen a year you have already signed.
4. Turn the agent off at the end of the job. Disconnect the company, or drop it to read-only, when the checklist is finished. Access that stays on for convenience becomes the access nobody remembers granting.
5. One client at a time, one task at a time. The whole product is built for that cadence; resist the temptation to run six clients' month-ends in parallel because it is possible.
6. Name the human who reviews each agent's work, per client, in writing. The most common failure in AI bookkeeping is not a bad model — it is a good model whose output nobody owned.

And one commercial rule: keep your checklists and client context exportable. They are the accumulated value of using this product, and they currently live in one vendor's workspace. Anything you would be sorry to lose should exist in a document you control.

Evidence & Test Log — checked October 22, 2026

Every number in this review is one of three things: something we read ourselves on Meridian's or Pilot's public pages, something Pilot states in its own launch material, or a comparison figure from another vendor's public page. This log separates them and dates the check. It also states plainly what we did not do.

Specs and limits that decide the purchase

What we did not test: we did not open a Meridian Starter account, did not connect any real QuickBooks or Xero ledger, did not add a connector to Claude, ChatGPT or Codex, and did not observe an agent read from or write to a live set of books. Every price, feature, security claim and scale figure in this review comes from Meridian's and Pilot's own pages and release, read on 22 October 2026, plus the public pages of Xero, Intuit, Financial Cents and Cube where a comparison figure is attributed to them. No number here is our own measurement of speed, accuracy, time saved or cost. Where official statements conflict — the free tier is the clearest case — we report both rather than choosing one.

Frequently Asked Questions