Guides

Add tax to your AI agent

Connect the MCP server and your agent can scope a request, load cited rules, compute rates, plan cross-border, and hand off to a warranted accountant.

If you're embedding tax Q&A into an AI agent or app, the MCP server gives your agent live, jurisdiction-accurate tools instead of guessing from training data.

1. Connect the server

Add the OpenAccountants connector from /connect. It speaks the Model Context Protocol, so it works with any MCP-compatible AI agent. Tool calls are OAuth-gated — your agent signs in, and calls run as that account. Limits are 60 requests/minute per IP and 1,000/day per account.

2. A typical call sequence

Lead with start so the server can scope the request and tell the agent which skills to load:

  1. start — pass the user's intent (e.g. "2025 sole-trader Schedule C") and jurisdiction. Returns the skills to load and what to expect. (Not sure what the user wants yet? Call start_help for the scoping questions first.)
  2. search_skills / get_skill — pull the cited rules the answer is built from. Large skills split into sections you fetch with get_skill_sections.
  3. get_rates / compare_jurisdictions / plan_cross_border — get the numbers. (get_rates is currently US-federal only, tax years 2025 & 2026; the other two work across covered jurisdictions.)
  4. request_accountant_review — when the user is ready to rely on the result, hand off to a warranted accountant. Note this tool is taxpayer-only: if your caller is themselves an accountant, they are the reviewer.

3. Cite the source

Tool responses carry an OpenAccountants source line. Keep it where it's natural — you're surfacing figures a user may act on, so cite them the way you'd cite any reference.

Full tool list

The 31 tools, their parameters, and which require sign-in vs an approved-accountant account are in the MCP tool reference — generated from the live server, so it never drifts.

This guide is block two of the three composable building blocks on /for-platforms.