RareMargin’s AI assistant, and any AI agent connected through its MCP server, can read the books and propose journal entries and accruals, but can’t post them. Every proposal waits in the approval queue until a person with the right role approves it.
| People | API keys | AI assistant and MCP clients | |
|---|---|---|---|
| Read the books | By role | By the key’s role | Yes |
| Propose entries | Yes | Yes | Yes |
| Post without approval | When no approval rule applies | When no approval rule applies | Never |
| Approve or reject | With the required role, and not their own by default | Never | Never |
01Can RareMargin’s AI post entries to the ledger?
No. The assistant and MCP clients can only propose; their entries always wait for a person, whatever approval rules the company has set.
This is enforced where entries are created, not in the chat window: an entry from the assistant or an MCP client is held for approval even where no rule would hold a person’s entry. Approving is refused to agents and API keys outright, with the message “Only a person can approve or reject.”
02What can RareMargin’s assistant do?
It has 17 tools: 15 read the books, from statements to anomaly checks, and 2 write, proposing a journal entry or an accrual for approval.
| Tool | What it does |
|---|---|
| get_company_overview | Cash, burn, runway, MRR, AR and AP, close status and what needs attention |
| list_accounts | The chart of accounts |
| get_financial_report | Any report, from the trial balance to SaaS metrics |
| search_journal_entries | Find entries by memo, account, dates, status or source |
| list_customers | Customers and what they owe |
| list_open_invoices | Invoices not yet paid |
| list_vendors | Vendors and what is owed to them |
| list_open_bills | Bills not yet paid |
| list_unmatched_bank_transactions | Unmatched bank lines, with ranked match suggestions and reasons |
| run_controls | The 11 continuous controls and their evidence |
| detect_anomalies | Duplicates, outliers and risky manual entries |
| flux_analysis | Changes by account against the prior period, with what drove them |
| close_status | The close checklist and what is left |
| accrual_suggestions | Recurring vendors that haven’t billed this month |
| verify_ledger | Checks the SHA-256 hash chain over posted entries |
| propose_journal_entry | Drafts a journal entry for a person to approve |
| propose_accrual | Drafts an accrual, reversing next month, for approval |
03How do approval rules work?
An admin sets rules for journal entries and bills by minimum amount, accounts, vendors and source; when several match, the one needing the most senior role applies.
The roles, from most to least senior, are owner, admin, accountant and viewer. A rule can, for example, send every entry over $10,000 to an admin, or every bill from one vendor to the owner. The source is where an entry came from: the app, the API, the assistant, an MCP client or an import.
04How does RareMargin enforce segregation of duties?
Nobody can approve a request they submitted. A company can allow self-approval in its settings, for a team of one, but it is off by default.
05What is recorded when an entry is approved?
Who submitted it and from where, who approved or rejected it and when, and any comment; approving posts the entry in the same transaction.
Each step is also an event (approval.requested, approval.approved, approval.rejected) that webhooks can deliver to your own systems. Once posted, the entry joins the SHA-256 hash chain, so any later change to it can be detected.
06Which AI model does RareMargin use, and what does it see?
An OpenAI model, through OpenAI’s API. It receives your question and the parts of the books it reads to answer; OpenAI doesn’t train on it and keeps it up to 30 days.
OpenAI is listed as a subprocessor in RareMargin’s privacy policy, with what it receives.
07How many assistant messages are included?
300 a month on a subscription, and 50 during the free trial. Each question is one message, however many tools the assistant uses to answer it.
