Build a legislation Q&A agent

This guide answers a plain-language question about a title in three calls: find the title, find the provisions about the question, fetch the text. The answer carries everything a reader needs to rely on it: the compilation, the status as confirmed with the Federal Register, the verification label and a citation they can follow.

Before you begin

Python 3.10 or later and the official SDK: pip install mcp.
A partner API key in LAWATLAS_API_KEY. See Get a credential.

Steps

Treat errors as answers. Every tool returns JSON. Check for error before reading anything else, and branch on error.code. A lapsed subscription returns plain text with a renewal link, so handle that too.
Find the title with resolve_reference, and stop if it is not in force: Advisor will not serve its text, and neither should your agent.
Find the provisions with search_provisions, scoped to the title and to sections.
Fetch the text with get_provision_text, and keep its status, anchor and link with it.
Show the status with the text. Say when a newer compilation exists (state other than current).

import asyncio
import json
import os

from mcp import ClientSession
from mcp.client.streamable_http import streamablehttp_client

URL = os.environ.get("LAWATLAS_MCP_URL", "https://advisor.lawatlas.ai/api/mcp-http/mcp")


class AdvisorError(Exception):
    def __init__(self, error):
        super().__init__(f"{error['code']}: {error['message']}")
        self.code = error["code"]
        self.error = error


async def call(session, tool, arguments):
    """Call one tool and return its JSON, raising on an error envelope."""
    result = await session.call_tool(tool, arguments)
    text = result.content[0].text
    try:
        body = json.loads(text)
    except ValueError:
        raise RuntimeError(text)  # plain text: a lapsed subscription's renewal message
    if isinstance(body, dict) and "error" in body:
        raise AdvisorError(body["error"])
    return body


async def answer(title, question):
    headers = {"Authorization": f"Bearer {os.environ['LAWATLAS_API_KEY']}"}
    async with streamablehttp_client(URL, headers=headers) as (read, write, _):
        async with ClientSession(read, write) as session:
            await session.initialize()

            # 1. Which title? The best candidate; show close alternatives to your user.
            found = await call(session, "resolve_reference", {"text": title})
            best = found["candidates"][0]
            if not best["in_force"]["in_force"]:
                return f"{best['title']} is not in force. See {best['in_force']['frl_link']}"

            # 2. Which provisions are about the question?
            hits = await call(session, "search_provisions", {
                "query": question, "legislation_id": best["id"],
                "provision_types": ["section"], "limit": 3})
            if not hits["result"]:
                return f"No provision of {best['title']} matched. Try other words."
            top = hits["result"][0]

            # 3. The text, with its status and verification.
            prov = await call(session, "get_provision_text", {
                "legislation_id": best["id"], "provision_id": top["provision_id"]})
            status, comp = prov["status"], prov["status"]["compilation"]
            lines = [
                f"{best['title']}, {prov['pinpoint']} {prov['heading']}",
                prov["text"],
                "",
                f"Compilation {comp['compilation_no']} ({comp['register_id']}), state: {comp['state']}.",
                f"In force, as confirmed with the Federal Register at {status['confirmed_with_frl_at']}.",
                f"Text check: {prov['anchor']['label']}.",
                f"Cite: https://advisor.lawatlas.ai{prov['link']['url']}" if prov.get("link") else
                f"Source: {status['frl_link']}",
            ]
            if comp["state"] != "current":
                lines.append("A newer compilation is registered; check the Federal Register before relying on this.")
            return "\n".join(lines)


if __name__ == "__main__":
    print(asyncio.run(answer("privacy act", "notifying individuals of an eligible data breach")))

Check it worked

Run python qa_agent.py. Run on 10 October 2026, it printed section 26WL of the Privacy Act 1988 ("Entity must notify eligible data breach") with these closing lines:

Compilation 104 (C2026C00227), state: current.
In force, as confirmed with the Federal Register at 2026-10-09T16:37:46.550767Z.
Text check: Verified.
Cite: https://advisor.lawatlas.ai/l/C2004A03712/compilation_104/main/sec_26wl

Letting a model choose the tools

To let a language model decide which tools to call, connect Advisor to an MCP client (Connect a client): the model then calls these same tools itself. Two rules keep its answers sound.

The law comes from the tool results, never from the model's memory. Instruct the model to quote only text a tool returned, and to say so when nothing was returned.
The status travels with the text. Instruct the model to state confirmed_with_frl_at, the compilation and the citation with every quotation, and to treat TITLE_NOT_IN_FORCE as an answer, not a gap to fill.

What to do if it did not