TaxChad speaks MCP (Model Context Protocol). That means Claude, ChatGPT, Cursor and any other MCP-compatible assistant can read a client's books, summarise unreconciled items or draft a payment link — on your instruction, using your login, under Row Level Security. No CSV exports. No shared passwords.
You already answer the same three questions ten times a week: what's outstanding, what's this transaction, how much is set aside for tax?An MCP-connected assistant answers them straight from live data, and drafts the follow-up — so you spend the hour on judgement, not on lookups.
"Which of my clients have unreconciled items older than 30 days?" — one prompt, one answer, no spreadsheet.
Payment links, chase emails, month-end summaries. Your assistant proposes, you approve.
Clients live inside Claude or ChatGPT. Give them a safe, read-only window into their own books, on your terms.
Log in once. RLS decides what the assistant can see — exactly the same rules that govern the app.
In Claude Desktop or Claude on the web: click the + in the composer → Connectors → Add connector → Add custom connector. In ChatGPT (Team/Enterprise/Pro): Settings → Connectors → Add. Cursor, Zed, Codex and Warp have the same option under Tools or MCP.
Any label works — TaxChad is the obvious one so it's easy to spot in your connector list later.
In the Remote MCP server URL field, paste https://taxchad.com/mcp. Leave Advanced settings alone — no OAuth Client ID or Secret to enter. Your assistant discovers everything it needs from TaxChad automatically. Click Add.
A TaxChad tab opens. Sign in with the Google account (or email) you use for TaxChad, then click Approve on the consent screen. You'll see a Connected page — close the tab and head back to your assistant.
Try: "whoami — what companies can you see in TaxChad?" Then move on to "show me last week's unreconciled bank transactions for [client]."
MCP looks like magic; the plumbing is deliberately boring. Here is exactly what happens — and what we refuse to do — every time an assistant calls TaxChad.
No API keys, no shared secrets. The assistant gets a short-lived token bound to your Supabase session and your consent. Revoke it any time from your account.
Every tool the assistant can call today is read-only against your bookkeeping data. It cannot delete transactions, edit invoices, or change client settings.
Requests run as you, through the same RLS policies as the web app. If you can't see it in TaxChad, neither can the assistant.
Tool, arguments and outcome are written to your MCP audit log. Open it any time under Settings → AI Access → Audit log.
Payouts, VRP collections and any tool that moves money are behind a platform kill switch, a per-company opt-in, and a two-step preview + confirmation-token flow. Compliance sign-off is pending, so today they refuse with `compliance_gate`.
TaxChad does not send your data to any model provider for training. What the assistant sees stays between you, the assistant vendor and TaxChad.
As a CIMA, ICAEW, ACCA or AAT-regulated professional you remain responsible for client confidentiality and for reviewing any output the assistant proposes. Treat MCP the same way you would a junior — helpful, fast, and always reviewed before you sign anything off.
Free to start. Add MCP whenever you're ready.