Kontor MCP
Enables AI assistants to locally parse, validate, audit, explain, generate, and convert XRechnung and ZUGFeRD/Factur-X e-invoices using official rule sets, fully offline with no API keys required.
README
Kontor MCP
The sovereign e-invoice toolkit for AI agents. Official XRechnung / EN 16931 validation, ZUGFeRD in and out, 100 % offline, zero API keys.
Kontor MCP is an open-source Model Context Protocol server that gives AI assistants (Claude Desktop, Claude Code, any MCP client) fully local capabilities for German/EU electronic invoicing: parse, validate, audit, explain, generate and convert XRechnung and ZUGFeRD/Factur-X invoices with the official KoSIT / EN 16931 rule sets — and no invoice data ever leaving your machine.
Status: v0.9 — full tool surface (8 tools, 4 resource families, 3 prompts), KoSIT conformance 89/89, ZUGFeRD PDF/A-3 generation verified by veraPDF and Mustang. Phase 3 (Streamable HTTP + auth, Docker, kontor-agent CLI, security pass, npm publish → v1.0) is in progress; until the npm publish, install from source as shown below.

Claude Desktop with the kontor server: validate → BR-DE-15 (missing Leitweg-ID) with fix hint · parse a ZUGFeRD PDF · explain BR-DE-18. MP4
Why
- Official rules, not approximations. The KoSIT XRechnung Schematron and the CEN EN 16931 rules run unmodified (compiled to XSLT/SEF, executed with Saxon-JS) with the KoSIT scenario model — the same verdicts the public-sector receivers produce, proven file-by-file against the official validator.
- Sovereign by construction. Pure TypeScript/WASM, no Java at runtime, no network calls, nothing stored, nothing logged. Runs where the invoices are.
- Agent-native. Every tool returns structured content and a readable summary; findings carry DE/EN explanations, affected business terms and fix hints; generation is fail-honest (
validis the real verdict, never assumed). - ZUGFeRD both ways. Read the embedded XML out of any ZUGFeRD/Factur-X PDF; write PDF/A-3 invoices with
factur-x.xmlthat pass veraPDF and Mustang.
Five-minute quickstart
Requires Node ≥ 20 and pnpm 10 (corepack enable picks the pinned version).
git clone https://github.com/DashankaNadeeshanDeSilva/kontor-mcp.git
cd kontor-mcp
pnpm install && pnpm build # ~1 min; all standards artefacts are bundled — no extra downloads, no Java
Claude Desktop — Settings → Developer → Edit Config, then quit (⌘Q) and reopen:
{
"mcpServers": {
"kontor": {
"command": "node",
"args": ["/absolute/path/to/kontor-mcp/packages/server/dist/bin.js"]
}
}
}
Claude Code:
claude mcp add kontor -- node /absolute/path/to/kontor-mcp/packages/server/dist/bin.js
Now audit a sample invoice. Ask Claude:
Audit
/absolute/path/to/kontor-mcp/packages/server/samples/broken-missing-buyer-reference.xml— is it valid?
You get a verdict (invalid), the finding BR-DE-15 (missing buyer reference / Leitweg-ID) with an explanation and fix hint, the recomputed totals and VAT breakdown, and a reject recommendation with its rationale. Then try a ZUGFeRD PDF:
What is in
/absolute/path/to/kontor-mcp/packages/server/samples/generated-zugferd-en16931.pdf?
(Attach XML files directly if you prefer; PDFs must be referenced by local path — Claude Desktop does not hand PDF bytes to MCP servers.)
Without a chat client, the MCP Inspector works headless:
npx @modelcontextprotocol/inspector@latest --cli node packages/server/dist/bin.js \
--method tools/call --tool-name audit_invoice \
--tool-arg file_path=$PWD/packages/server/samples/broken-missing-buyer-reference.xml --tool-arg lang=en
kontor-agent CLI — the reference client; audit needs no LLM and returns 0/1/2 for accept/review/reject, chat is an Anthropic agent loop that prints every tool call (needs ANTHROPIC_API_KEY):
node packages/client/dist/bin.js audit packages/server/samples/broken-missing-buyer-reference.xml --lang en
node packages/client/dist/bin.js chat -m "Prüfe $PWD/packages/server/samples/valid-zugferd-en16931.pdf"
node packages/client/dist/bin.js --url http://127.0.0.1:3333/mcp --token "$KONTOR_AUTH_TOKEN" tools # against Docker / HTTP
Docker / Streamable HTTP — the same server over HTTP for remote agents and containers (token required, loopback-published port, TLS is your reverse proxy's job):
docker build -t kontor-mcp . # multi-stage, non-root, ~70 MB, amd64 + arm64
docker run -d -p 127.0.0.1:3333:3333 -e KONTOR_AUTH_TOKEN="$(openssl rand -hex 24)" kontor-mcp
curl -s http://127.0.0.1:3333/healthz # {"ok":true,"name":"kontor-mcp",...}
Or cp .env.example .env && docker compose up -d (read-only root FS, ./invoices mounted at /data). Config surface and the security posture: packages/server/README.md, SECURITY.md.
Tools
| Tool | What it does |
|---|---|
parse_invoice |
Detect the format (UBL / CII · EN 16931 · XRechnung version & variant · ZUGFeRD profile) and return the EN 16931 semantic model |
validate_invoice |
XSD + official EN 16931 / XRechnung Schematron (KoSIT scenarios) + Kontor plausibility → valid / valid_with_warnings / invalid with explained findings |
audit_invoice |
One call for accounts payable: header facts, VAT breakdown, verdict, grouped findings, accept / review / reject with rationale; stateless duplicate detection via known_invoice_numbers |
generate_invoice |
Structured data → XRechnung 3.0 (UBL) or ZUGFeRD 2.3 / Factur-X PDF/A-3 (target: zugferd-pdf, profiles EN16931 / BASIC / EXTENDED); decimal-safe amounts, internal validation, deterministic auto-fixes reported |
convert_invoice |
ZUGFeRD PDF → XML, UBL ↔ CII via the semantic model (post-validated, honest loss report), self-contained HTML preview |
check_obligations |
German e-invoicing mandate decision tree (B2B/B2G/B2C, 2025 → 2028 transition, exemptions) with primary legal sources |
explain_rule |
Official text, explanation, affected business terms and fix hint for any BR-* / BR-DE-* / KONTOR-* rule id |
list_capabilities |
Formats, bundled standard versions, KB stats, legal lastVerified, inventory, sovereignty statement |
Resources: kontor://samples/{name}, kontor://reference/rules, kontor://reference/codelists/{list}, kontor://reference/cheatsheet. Prompts: audit-incoming-invoice, draft-supplier-rejection, create-invoice-interview. Full reference with inputs and conventions: packages/server/README.md.
Sovereignty — and how we prove it
| Claim | Proof |
|---|---|
| No network at runtime | No HTTP client in the dependency tree; every rule set, schema, code list and legal fact is bundled in @kontor-mcp/rules with provenance and checksums. A network-guard test is part of the Phase 3 security pass. |
| No Java, no native code | Schematron is compiled at build time and executed with Saxon-JS; XSD via xmllint-wasm; PDF via pdf-lib. Java is used only by the development-time oracles (KoSIT validator, veraPDF, Mustang) in CI. |
| Nothing stored, nothing logged | The server is stateless; invoice contents never reach a log (KONTOR_LOG_PAYLOADS defaults to off). |
| The verdicts are the official ones | docs/CONFORMANCE.md: 89/89 files of the official XRechnung test suite with identical verdicts and identical findings vs the KoSIT validator, replayed on every CI run. |
| The PDFs are real PDF/A-3 | Every generated sample is regenerated in CI and checked by veraPDF (PDF/A-3b, zero violations) and Mustang CLI (PDF/A + XMP + profile XSD + EN 16931 rules). |
Architecture
flowchart LR
subgraph client [MCP client]
A[Claude Desktop / Claude Code / kontor-agent]
end
A -- stdio / Streamable HTTP --> S
subgraph server ["@kontor-mcp/server"]
S[tools · resources · prompts<br/>Zod schemas, structuredContent + text]
end
subgraph core ["@kontor-mcp/core (MCP-free library)"]
D[detect] --> P[parse → EN 16931 model]
P --> V[validate: XSD → Schematron → plausibility]
P --> G[generate / convert / preview]
G --> Z[ZUGFeRD PDF/A-3 assembly]
V --> AU[audit → verdict + recommendation]
O[obligations decision tree]
end
subgraph rules ["@kontor-mcp/rules (bundled, checksummed)"]
R1[XSDs UBL 2.1 / CII D16B]
R2[EN 16931 + XRechnung Schematron as SEF]
R3[rule knowledge base DE/EN]
R4[code lists · legal timeline · fonts + ICC]
end
S --> core
core --> rules
The validation pipeline: Layer 1 XSD (xmllint-wasm) → Layer 2 official Schematron (Saxon-JS, KoSIT scenario selection and severity overrides) → Layer 3 Kontor plausibility (KONTOR-PLAUS-*: decimal recomputation, VAT rates, IBAN/BIC, Leitweg-ID check digits, dates, duplicates — never changes the official verdict). Money math is decimal.js only; every XML parse has DTD/external entities disabled.
Conformance
| What | Reference | Result (2026-08-25) |
|---|---|---|
| Validation verdicts and findings | KoSIT validator 1.6.3, XRechnung 3.0.2 test suite (86 files) + Kontor fixtures (3) | 89/89 verdict and finding parity |
| Generated ZUGFeRD PDFs | veraPDF 1.30.2 (PDF/A-3b) | 6/6 PASS, zero violations |
| Generated ZUGFeRD PDFs | Mustang CLI 2.26.0 (PDF/A + XMP + Factur-X profile XSD + EN 16931) | 6/6 valid (EN 16931, BASIC, EXTENDED × DE/EN) |
| Generated XRechnung | own pipeline (identical to the oracle above), 50 random inputs | 50/50 valid and plausible |
Details, commands and recorded reports: docs/CONFORMANCE.md.
Packages
| Package | Purpose |
|---|---|
@kontor-mcp/core |
MCP-free library: detect, parse, validate, audit, generate, convert, ZUGFeRD PDF |
@kontor-mcp/rules |
Bundled standards artefacts (XSDs, compiled Schematron, code lists, legal timeline, PDF assets) + rule knowledge base |
@kontor-mcp/server |
MCP server (stdio and Streamable HTTP with bearer auth; Docker image) exposing tools, resources, prompts |
@kontor-mcp/client |
kontor-agent — reference MCP client CLI: tools introspection, scriptable audit <file> (no LLM, exit codes), chat Anthropic agent loop with tool-call trace; stdio or HTTP |
FAQ
Is this legal or tax advice? No. Kontor reports formal and technical checks against the published standards and the legal timeline with its sources; decisions remain yours. Every answer carries that disclaimer.
Does it support XRechnung 3.0.2 / the 2025 rules? Yes: XRechnung 3.0.2, Schematron 2.5.0, validator configuration 2026-01-31, EN 16931 1.3.16. Versions are pinned and listed by list_capabilities.
Which ZUGFeRD profiles can it read and write? Read: MINIMUM, BASIC WL, BASIC, EN 16931, EXTENDED, XRECHNUNG (profile from the XMP). Write: EN 16931 (default), BASIC, EXTENDED — see D-042 for what BASIC drops.
Can I send it a PDF from Claude Desktop? Reference it by local path. Desktop does not pass attached PDF bytes to MCP servers; XML attachments work either way.
Why no Java when KoSIT's validator is Java? Sovereignty and installability: Schematron is compiled once at build time and executed in TypeScript (Saxon-JS). The Java tools are used only as oracles in CI to prove parity.
Windows? Yes — CI runs the full test suite on Ubuntu, macOS and Windows with Node 20 and 22.
Development
pnpm install
pnpm build
pnpm -r test
pnpm lint
pnpm artifacts # fetch pinned third-party artefacts for the oracles (dev only)
pnpm oracle --diff … # KoSIT parity (Java 17+)
pnpm check:zugferd # regenerate samples, veraPDF + Mustang (Java + veraPDF)
Read CONTRIBUTING.md before opening a PR; security issues go through SECURITY.md.
Docs
- Product Requirements · Implementation Plan · Decision log
- Conformance · Business-term coverage · Artefact provenance
- v0.1 Claude Desktop verification
License
Apache-2.0 — see LICENSE and NOTICE for bundled third-party components (KoSIT artefacts Apache-2.0, EN 16931 artefacts EUPL-1.2, Liberation Fonts OFL 1.1, ICC sRGB profile).
推荐服务器
Baidu Map
百度地图核心API现已全面兼容MCP协议,是国内首家兼容MCP协议的地图服务商。
Playwright MCP Server
一个模型上下文协议服务器,它使大型语言模型能够通过结构化的可访问性快照与网页进行交互,而无需视觉模型或屏幕截图。
Magic Component Platform (MCP)
一个由人工智能驱动的工具,可以从自然语言描述生成现代化的用户界面组件,并与流行的集成开发环境(IDE)集成,从而简化用户界面开发流程。
Audiense Insights MCP Server
通过模型上下文协议启用与 Audiense Insights 账户的交互,从而促进营销洞察和受众数据的提取和分析,包括人口统计信息、行为和影响者互动。
VeyraX
一个单一的 MCP 工具,连接你所有喜爱的工具:Gmail、日历以及其他 40 多个工具。
graphlit-mcp-server
模型上下文协议 (MCP) 服务器实现了 MCP 客户端与 Graphlit 服务之间的集成。 除了网络爬取之外,还可以将任何内容(从 Slack 到 Gmail 再到播客订阅源)导入到 Graphlit 项目中,然后从 MCP 客户端检索相关内容。
Kagi MCP Server
一个 MCP 服务器,集成了 Kagi 搜索功能和 Claude AI,使 Claude 能够在回答需要最新信息的问题时执行实时网络搜索。
e2b-mcp-server
使用 MCP 通过 e2b 运行代码。
Neon MCP Server
用于与 Neon 管理 API 和数据库交互的 MCP 服务器
Exa MCP Server
模型上下文协议(MCP)服务器允许像 Claude 这样的 AI 助手使用 Exa AI 搜索 API 进行网络搜索。这种设置允许 AI 模型以安全和受控的方式获取实时的网络信息。