会計士 監査調書 scaffold pack (A2)

被監査会社 360 度 (法人格 × 適格事業者 × 行政処分 × 採択履歴 × 法改正) を 1 packet。研究開発税制 + IT 導入補助金 会計処理 dual check + 監査調書 scaffold 自動生成。意見表明は会計士責任の境界を維持。

誰のためか

会計士 (CPA-Firm) コホート. 月次 300 req 想定 (¥3/req)。

1 query で返るもの (package_kind)

evidence_packet, artifact_pack

呼び出される heavy / composed endpoint

• POST /v1/cases/cohort_match
• GET /v1/tax_rules/{rule_id}/full_chain
• POST /v1/dd/question_match
• POST /v1/houjin/360
• POST /v1/evidence/packets/query

SDK example — Python (jpcite-mcp / autonomath-mcp PyPI)

# A2 CPA Audit Workpaper Scaffold Pack — Python (jpcite-mcp)
import os
import httpx

JPCITE_API = "https://api.jpcite.com"
API_KEY = os.environ["JPCITE_API_KEY"]  # see https://jpcite.com/dashboard

def fetch_kaikei_packet(houjin_bangou: str) -> dict:
    # 1. Free preflight — no charge until accepted_artifact delivery.
    preview = httpx.post(
        f"{JPCITE_API}/v1/jpcite/preview_cost",
        headers={"X-API-Key": API_KEY, "X-Client-Tag": "kaikei-001"},
        json={"outcome_contract_id": "A2_cpa_audit_workpaper_pack",
               "input": {"houjin_bangou": houjin_bangou}}
    ).raise_for_status().json()
    assert preview["cap_passed"], "outcome exceeds cap — adjust max_price_jpy"

    # 2. Execute packet — charges 75 billable units (=¥75).
    packet = httpx.post(
        f"{JPCITE_API}/v1/jpcite/execute_packet",
        headers={"X-API-Key": API_KEY,
                  "X-Client-Tag": "kaikei-001",
                  "X-Jpcite-Scoped-Cap-Token": preview["scoped_cap_token"],
                  "Idempotency-Key": f"kaikei-{houjin_bangou}-2026-05"},
        json={"outcome_contract_id": "A2_cpa_audit_workpaper_pack",
               "input": {"houjin_bangou": houjin_bangou},
               "max_price_jpy": 2500}
    ).raise_for_status().json()
    return packet  # contains source_url[], fetched_at, known_gaps[], evidence[]

SDK example — TypeScript (Bedrock MCP / OpenAI tool-use)

// A2 CPA Audit Workpaper Scaffold Pack — TypeScript (Bedrock / OpenAI tool-use)
import { JpciteClient } from "@jpcite/sdk";  // wraps OpenAPI 219 paths

const client = new JpciteClient({
  baseUrl: "https://api.jpcite.com",
  apiKey: process.env.JPCITE_API_KEY!,
  clientTag: "kaikei-001",  // 顧問先 fan-out attribution
});

export async function fetchCPAFirmPacket(houjinBangou: string) {
  const preview = await client.previewCost({
    outcome_contract_id: "A2_cpa_audit_workpaper_pack",
    input: { houjin_bangou: houjinBangou },
  });
  if (!preview.cap_passed) throw new Error("over cap");

  const packet = await client.executePacket({
    outcome_contract_id: "A2_cpa_audit_workpaper_pack",
    input: { houjin_bangou: houjinBangou },
    max_price_jpy: 2500,
    scopedCapToken: preview.scoped_cap_token,
    idempotencyKey: `kaikei-${houjinBangou}-2026-05`,
  });
  return packet;  // { source_url[], fetched_at, known_gaps[], evidence[] }
}

Claude Code .mcp.json snippet

{
  "mcpServers": {
    "jpcite": {
      "command": "npx",
      "args": [
        "-y",
        "autonomath-mcp"
      ],
      "env": {
        "JPCITE_API_KEY": "${env:JPCITE_API_KEY}",
        "JPCITE_DEFAULT_OUTCOME": "A2_cpa_audit_workpaper_pack",
        "JPCITE_CLIENT_TAG": "kaikei-001"
      }
    }
  }
}

jpcite ¥900 / 月 vs LLM API replacement

cheapest LLM tier (Haiku 4.5) でも jpcite の 0.5x コスト。 さらに LLM 単独は出典 URL を hallucinate するため、 一次資料 URL + fetched_at + known_gaps は jpcite でしか得られない。

コホート 別コスト比較表を見る →

1 call ≈ 人手 ¥2,500 相当

30-90 分の一次資料調査 + 名寄せ + 出典確認を 1 query で代替。 代替時給 ¥3,000 換算で 1 packet ≒ ¥2,500。 jpcite ¥3 / packet との差分が ROI。

使い始める

1. /dashboard で API キー発行 (匿名 3 req/日無料)
2. Python: pip install autonomath-mcp
3. Claude Code: .mcp.json 上記 snippet を貼り付け
4. getting-started 経由で 1 packet を取得

関連リソース

• API reference (219 path)
• MCP tools (146 default-gate)
• cookbook (15 recipes)
• 会計士 向け コホート landing
• outcome catalog (450 outcome)