Monday morning, client A. You open Claude, paste last month's P&L export, explain the sign convention, remind it that the fiscal year ends in June, and get a good variance summary in ten minutes. Tuesday, client B: same ritual, different ledger, different fiscal year. Thursday, back to client A, and Claude has no idea what you agreed on Monday.
Claude is not the problem. The setup is a chat, and a chat is built for one conversation. A fractional practice is four or five ongoing engagements, each with its own books, rules and history. This guide turns Claude from a chat you re-brief every day into a setup that remembers each client: what to configure, in what order, and what to keep out.
It is the practical companion to AI for Fractional CFOs, which covers what to hand to AI and what to sign. The principle underneath both is what we call the CFO harness: the layer around the model that makes its work signable.
What Does a Multi-Client Claude Setup Look Like?
Five layers, each answering a question you otherwise answer by hand:
| Layer | Question it answers | Where it lives |
|---|---|---|
| Account | Whose subscription, whose data terms? | Your plan or the client's |
| Project per client | Which client am I working on? | A Claude Project |
| Connectors | Where do the actuals come from? | MCP connections, read-only |
| Persistent model + FINANCE.md | What is the model, and how do I read it? | A structured model layer |
| Procedures | How do we run the close, the pack, the commentary? | Skills or written checklists |
Set them up in that order. Each one is useless without the one before it.
Step 1: Whose Account Does Each Client Run On?
Decide this before anything else, because it decides what you can connect. There are two clean options.
The client's account. The client has a Claude Team or Enterprise plan and adds you as a member. Their data terms apply, their admin controls the connectors, and when you leave, the work stays with them. This is the cleanest answer for a client with strict data rules.
Your account, with the client's consent. Most small clients do not have an AI plan. You work in your own (Pro, Max or Team), with one Project per client, and the client agrees in writing to which data goes in and on what terms. What those clauses usually cover is in Your Client's Contract Says No AI, and what happens to data on each plan is in AI and Confidential Financial Data.
What does not hold up is the middle: client data in a personal account nobody agreed to, mixed in with other clients.
Step 2: How Should You Use Claude Projects for Each Client?
One Project per client, never shared between clients. A Project groups the chats, the reference files and the instructions for one engagement, so every new chat starts inside that client's context rather than an empty one.
Keep the Project instructions short. They should say who the client is and where things live, not restate every convention:
Client: Distrib SAS, food distribution, 40 employees, France.
Fiscal year: July to June. Reporting currency: EUR, in thousands.
Materiality for commentary: variances above 5% and above 10k.
The financial model lives in Layerz (model "Distrib 2027 budget").
Read its FINANCE.md before answering any question about figures.
Never post or edit anything in the client's ledger.
Reference files that belong in a Project: the chart of accounts, the last board pack, the engagement scope. Files that do not: anything you would not want in the next chat by default, such as payroll detail or another client's template filled with that client's numbers.
The conventions themselves (sign, units, definitions, how the statements connect) are better kept with the model than in the Project. Instructions in a Project are read by Claude in that Project. Conventions stored with the model are read by any session, any assistant and any colleague who opens it. That is the difference between remembering a conversation and remembering a model, covered in Persistent Memory for a Financial Model.
Step 3: Which Connectors Should a Fractional CFO Use?
Connect the sources rather than pasting them, and connect them read-only. Claude talks to other systems over MCP (Model Context Protocol), an open standard. Three kinds of connection matter.
The ledger. Xero and Intuit both publish official MCP servers, so Claude can query a client's books directly. Those servers expose write operations as well as reads. On a ledger you do not own, read-only is the right scope: you analyse, the bookkeeper posts. Where there is no server you trust, export the reports (trial balance, general ledger, aged receivables) and load them; that is still how most fractional work starts. The fractional CFO tech stack lists what connects to what.
The bank and billing. Some banks and billing tools publish MCP servers or have native connectors in the tools you use. Same rule: read-only unless there is a reason.
The model. This is the connection most setups miss, and the next step.
Custom connectors over remote MCP are available on Claude Free, Pro, Max, Team and Enterprise, according to Anthropic's help center; Free accounts are limited to one. On Team and Enterprise, an owner adds the connector to the organisation and each member then connects. Anthropic's own warning is worth applying literally: custom connectors "have not been verified by Anthropic", so connect only servers you trust and read the permissions you grant at sign-in.
Step 4: Where Does Each Client's Model Live Between Sessions?
Outside the chat, on a layer Claude reads and writes over MCP. This is what turns Thursday's session into a continuation of Monday's.
A spreadsheet uploaded to a chat is a snapshot. Claude reads it, works on a copy, and the result is a new file. Next session, you upload again and the structure has to be re-understood. A persistent model is different: Claude opens the same model each time, sees its structure (items, formulas, timelines, how the P&L, cash flow and balance sheet connect), loads new actuals into it, and leaves it in a state the next session finds.
Give each client's model its own FINANCE.md: currency, units, sign convention, fiscal year, glossary, what the model is for, how the statements are built. Claude reads it before touching anything. We published FINANCE.md as an open standard for exactly this file; start with three lines (units, sign, fiscal year end) and add a definition each time a term gets misread.
Two properties matter more than the tool you pick:
- Claude does not do the arithmetic. Totals, balances and forecasts come from formulas evaluated by an engine. Claude proposes the formula and explains the result. A figure you sign should never be one Claude computed in its head.
- Every change is recorded. When Claude edits the model, you can see what changed, compare with last month and restore an earlier state.
Step 5: How Do You Make Claude Run the Same Procedure at Every Client?
Write the procedure down once and have Claude follow the written version, not your memory of last month's prompt. Claude supports Skills: a folder with instructions (and, if useful, scripts) that Claude loads when the task matches. A close-acceptance skill for a fractional practice might read:
- Read the client's FINANCE.md.
- Load this month's actuals into the model.
- Run the checks: balance sheet balances, cash closes into the balance sheet, no unmapped accounts.
- List variances above the client's materiality threshold, ranked.
- Draft the commentary for each, citing the model's figures.
- List the decisions that need the CFO, and stop.
Step 6 is the important one. The skill ends with your decisions, not with a sent email.
The same procedure runs at every client because the client-specific parts (threshold, fiscal year, definitions) come from that client's FINANCE.md and model, not from the skill. "Reuse the workflow, never the sign-off" is the rule we use for the month-end close across multiple clients. If you keep your procedures in a repository with history, GitHub for Fractional CFOs shows a one-repo-for-the-method, one-repo-per-client layout.
Fractional CFOs who build tools often move to Claude Code at this point, to script imports and checks. It is useful and not required. Everything above works in the Claude app.
What Should You Keep Out of Claude?
Five things, whatever your plan.
- One client's data in another client's Project. Not even as an example. Use a blank template for examples.
- Write access to ledgers you do not own. Analysis needs reads. Posting is the bookkeeper's job, with their review.
- Figures you sign that came from chat arithmetic. If a number goes into a pack, it comes from the model, with a formula behind it.
- Credentials. No passwords or API keys pasted into a chat. Connectors handle authentication.
- Client data in your personal account when the contract says otherwise. Check the engagement letter first, then ask. Some clients ban AI on their data, others require consent or specific retention terms.
A Weekly Routine That Holds
Start of a session, per client. Open the client's Project. Ask Claude to read the model's conventions and summarise what changed since your last session (the version history answers that). Thirty seconds instead of ten minutes of re-briefing.
During. Ask for changes to the model, not for numbers in the chat. Ask Claude to preview a change before applying it when it touches structure.
End. Name the version you would sign ("March close accepted"), so next month's comparison has a fixed point.
Monthly test. Open a fresh chat on a client you have not touched in two weeks and ask three questions: what is this model for, what feeds gross margin, what changed last. If the answers come from the setup rather than from you, it holds.
Where Layerz Fits
Layerz is the persistent model in Step 4. It is a structured spreadsheet, ready for AI: the structure of each model is kept separate from its data, so you reuse one structure across clients while the values stay in each client's own model.
Claude drives it over MCP. Add it as a custom connector with the URL https://layerz.cc/mcp and sign in (OAuth, no key to paste). Claude reads each model's FINANCE.md before acting, loads actuals from Pennylane, Qonto and Stripe through native connectors or from FEC, DATEV, Excel and CSV exports, builds scenarios as variants of the base case, and changes the model without breaking its formulas. Every change is versioned with a diff and restore (24 hours plus one pinned version on the free plan, 30 days on Team). For a stricter boundary, an API key can be bound to a single model and set to read-only. Every model exports to clean Excel with formulas intact, so the client keeps a file their next CFO can open.
It is free for one person working alone, with unlimited models. That is the shape of most fractional practices.
FAQ
Can I use one Claude account for all my clients? Yes, with one Project per client and the client's written consent where their contract requires it. For clients with strict data rules, working inside their own Claude Team or Enterprise plan is cleaner.
Do Claude Projects remember everything about a client? They keep that client's chats, files and instructions together, which removes most of the re-briefing. They do not hold a living financial model. For that, connect a persistent model layer and keep the conventions with the model.
Is it safe to connect Claude to a client's QuickBooks or Xero? It can be, with read-only scope, on an access the client has granted, and through an official server. Avoid giving an assistant write access to books you do not own.
Do I need Claude Code? No. Projects, connectors and Skills cover a multi-client setup in the Claude app. Claude Code helps if you want to script imports or checks.
Does this setup work with ChatGPT too? The principle does. The harness (persistent model, conventions, permissions, history) is independent of the assistant. We covered the ChatGPT side in ChatGPT for Fractional CFOs.
The Takeaway
Claude is very good at the language work around the numbers. What it lacks out of the box is a memory of each client, a boundary around each client's data, and a place for the numbers to live between sessions. That is all setup, and you do it once.
One account decision. One Project per client. Read-only connectors. A persistent model with its own FINANCE.md. Procedures written down. Then judge the setup by the only question that matters in a fractional practice: is the next client cheaper than the last?