Repository avatar
Other Tools
v0.5.0
active

beliq-mcp

io.github.beliq-eu/beliq-mcp

Validate, parse, generate and convert EU e-invoices (XRechnung, ZUGFeRD, Factur-X, Peppol BIS).

Documentation

beliq-mcp

An MCP server for the beliq e-invoice API. It lets MCP clients (Claude Desktop, Claude Code, Cursor, and others) validate, parse, generate, and convert electronic invoices (XRechnung, ZUGFeRD, Factur-X, Peppol BIS, and the other formats the API carries), and explain exactly what fails. Every verdict carries the check depth the API states for it, or says that none was stated.

beliq covers three steps of an e-invoice: generate, validate, and deliver (in development). This server runs the first two and also parses and converts documents; no tool sends or files an invoice. Until delivery through beliq is available, transmission (Peppol, plateforme agréée, SDI) and tax-authority reporting stay with your access point. Archiving stays with you, and beliq does not operate KSeF submission.

Tools

  • beliq_validate_einvoice - validate an invoice XML (inline or by file path) or a Factur-X/ZUGFeRD PDF (by file path). Returns the verdict, how deep the check went, the detected format and profile, the ruleset versions it was checked against, and every error and warning with its rule id, severity, location, and message. For a PDF the verdict is about the embedded XML; the PDF itself is not checked.
  • beliq_parse_einvoice - parse a UBL/CII XML invoice or a Factur-X/ZUGFeRD PDF into a structured EN 16931 invoice (fields such as number, dates, currency, seller, buyer, lines and totals). Returns the detected format and profile, the extracted invoice, and warnings that say what the document holds and the parsed invoice does not.
  • beliq_generate_einvoice - generate a document from an EN 16931 invoice object. XML comes back inline; a PDF is written to the outputPath you give. By default (verify) the API validates the document before returning it, and a document that fails comes back as a tool error with the failed rules. The tool description tells the model to ask for missing invoice data instead of making it up.
  • beliq_convert_einvoice - convert a document from one EN 16931 format to another. An XML source comes back as XML, inline, for every target, facturx and zugferd included. A hybrid PDF source sent to facturx or zugferd comes back as a PDF, written to outputPath: the PDF as sent, its container not checked, when its embedded XML already is the target; otherwise that PDF re-wrapped around the converted XML and passed by veraPDF against PDF/A-3b. With a test key the PDF is watermarked after any veraPDF check. The convert reference has the detail. Reports how many source elements the engine recorded as having no target equivalent.
  • beliq_check_account - verify the configured API key and report the plan, the remaining quota, and whether the key is live or a blq_test_ sandbox key, whose quota is the sandbox allowance. Calls GET /v1/me, which draws no quota; useful as a connection and credential smoke test.

Installation

Requires Node.js >= 20.15. Published to npm, so clients can run it with npx:

npx -y beliq-mcp

It is listed in the official MCP Registry as io.github.beliq-eu/beliq-mcp, which points clients at this npm package.

The server is configured entirely through environment variables (see below).

Configuration

VariableRequiredDefaultDescription
BELIQ_API_KEYyes-API key from the beliq dashboard (API Keys).
BELIQ_AUTHnoheaderHow the key is sent: header (X-API-Key) or bearer (Authorization: Bearer).
BELIQ_BASE_URLnohttps://api.beliq.euThe base URL the server sends its API requests to.

Client setup

Claude Code

claude mcp add beliq -e BELIQ_API_KEY=your-key -- npx -y beliq-mcp

Claude Desktop

Add to claude_desktop_config.json (Settings > Developer > Edit Config):

{
  "mcpServers": {
    "beliq": {
      "command": "npx",
      "args": ["-y", "beliq-mcp"],
      "env": {
        "BELIQ_API_KEY": "your-key"
      }
    }
  }
}

Cursor

Add to ~/.cursor/mcp.json (or a project .cursor/mcp.json) using the same mcpServers block shown for Claude Desktop.

Reading a result

beliq_validate_einvoice returns a short text verdict plus a structured result:

  • valid is the API's verdict: the document passed the rule packs and schemas beliq runs for its format, in the versions this result names. It does not mean every recipient, validation tool or tax authority accepts it. Warnings do not make a document invalid, with one case to read beside it: franceCtcBlockingRuleIds.
  • verificationBadgeLabel is the check depth as English display text, and the text verdict prints it beside VALID or INVALID. How verification works explains each depth. Where the API sends no label, the text says check depth not stated and names no depth. verificationBadge and verificationTier are the ids behind the label.
  • franceCtcBlockingRuleIds lists the French CTC rule ids the API reports as ending a transmission through a French platform. It can be present where valid is true, and the text verdict then leads with it. The result stays a normal result, not a tool error. franceCtcApplied, franceCtcVersion and franceExtendedCtcApplied say which French packs ran.
  • pdfInput is present when the document was a PDF: attachmentName, attachmentFoundBy and attachmentCandidates say which embedded file the verdict is about, and containerChecked is always false because the PDF itself is not checked.
  • format and profileDetected report the detected syntax and business profile.
  • schematronVersion is the ruleset revision the check ran against. The versions of the other packs and schema bundles that ran come in ciusVersion, peppolVersion, facturxProfile, facturxVersion, italyFatturapaXsdBundle, italyFatturapaRuntimeVersion, italySdiMessaggiXsdBundle, italySdiMessaggiXsdVersion, spainFacturaeXsdBundle, spainFacturaeRuntimeVersion, sloveniaEslogXsdBundle, sloveniaEslogRuntimeVersion, polandKsefFa3XsdBundle, polandKsefFa3RuntimeVersion, romaniaRoCiusVersion and netherlandsNlciusVersion, each present only when it applies.
  • rulesetChannel is present when a ruleset channel was selected, which an organisation whose default channel is previous does with no input of this tool. rulesetFellBack, rulesetSunsetsOn and rulesetSupersededOn come with it when they apply. rulesetSupersededOn is the date the authority made the successor of a selected ruleset mandatory, and a RULESET_SUPERSEDED warning names each such ruleset.
  • sha256 is the hash of the document that was checked, rulesetSha256 and rulesetArtifacts fingerprint the rule files that ran, and driftCheckedAt is when the nightly drift check last confirmed those files against their sources: each file's pinned download or, when the file cannot be fetched from there, a mirror the pin names.
  • verified is false only when no ruleset ran, which validate never returns; see generate below.
  • errors and warnings each carry ruleId, severity, location (an XPath when available), and message. errorCount and warningCount are their lengths.
  • livemode is false for a blq_test_ sandbox key and true for any other key. A test key runs the same rules and gives the same verdict as a live key.

The structured result carries every field of the API's validation result. Its shape is checked against the SDK's type, so a field the API adds cannot go missing here without failing the build.

beliq_parse_einvoice returns a text summary plus a structured result:

  • format, profileDetected and profileUrn report the detected syntax and profile; businessProcessId and franceCtcDetected are present when the document states them.
  • invoice is the extracted invoice object, with EN 16931 fields such as number, dates, currency, seller, buyer, lines and totals. The parse reference lists every field. A field the API holds no value for is left out.
  • warnings says what the document holds and the parsed invoice does not: each entry has a code, a message, and, by code, a field, the terms (EN 16931 business groups) or the elements (paths with counts) concerned. The text summary lists them.

beliq_generate_einvoice returns a short text summary (with the XML document appended for XML output) plus a structured result:

  • output is xml or pdf, as requested, and contentType is the matching media type (application/xml or application/pdf).
  • xml is the generated document inline, present only for XML output.
  • outputPath and bytesWritten are set when the document was written to disk: always for a PDF, and for XML when you set outputPath. The call never overwrites an existing file, so pick a path that does not exist. A path that is taken, or whose directory is missing or not writable, is refused before the API is called, so it costs no document.
  • pdfKind is present only for PDF output: hybrid (a PDF/A-3 with the XML embedded, for facturx and zugferd) or visualization (rendered pages with no XML inside, for xrechnung and peppol-bis, whose legal document stays the XML).
  • schematronVersion is the ruleset (Schematron) revision the document was checked against.
  • outputEnvelope is the envelope the document is wrapped in, when the API names one.
  • sha256 is the lowercase-hex SHA-256 of the returned document bytes, so sha256sum on the file at outputPath reproduces it.
  • rulesetSha256 is one combined fingerprint of the rule artifacts the document was checked against, present when a ruleset ran.
  • livemode is true for a blq_live_ key and false for a blq_test_ sandbox key, whose output is marked as a specimen and is not a production invoice. The text summary says so for a test key.
  • validationResult is the API's verdict on the generated document, with every field validate returns, the check depth in verificationBadgeLabel included; the text summary names that depth. The verdict is about the invoice XML, so for a PDF the text says so and names the kind of PDF. With the default verify: true, a document that fails validation comes back as a tool error that lists the failed rules. With verify: false no ruleset runs: verified is false, errors and warnings are empty because nothing checked the document, and the text says not validated.

beliq_convert_einvoice returns a short text summary (with the XML document appended for an XML target) plus a structured result:

  • output is what came back, read from the response: xml for every target when the source is XML, facturx and zugferd included, and pdf when a hybrid PDF source went to a facturx or zugferd target. contentType is the matching media type.
  • xml is the converted document inline, present only when XML came back.
  • outputPath and bytesWritten are set when the document was written to disk: always for a PDF, and for XML when you set outputPath. Like generate, it never overwrites an existing file, and it refuses a path it cannot write before the API is called.
  • sourceFormat is the format the engine read, and targetFormat the format it produced (the targetFormat you asked for when the API does not name one).
  • profileDetected is the profile the engine recognised on the source document, when it recognised one.
  • conversionTools names the toolchain that did the conversion, when the API names one.
  • livemode is true for a blq_live_ key and false for a blq_test_ sandbox key, whose output is marked as a specimen like generate's. The text summary says so for a test key.
  • lostElementsCount and lostElements count and list the source elements the engine recorded as having no target equivalent. A loss the engine does not detect is not counted, so 0 does not say that every field of the source arrived.
  • Each lostElements entry has the path, value, reason, severity and acceptableLoss that the LostElement schema describes. When the list would pass the API's 6 KB header cap, it holds one { "truncated": true } entry instead of the losses, and lostElementsCount stays the exact count.

A successful convert means the API validated the converted XML before returning it. For a facturx or zugferd target at profile en16931 that validation is the one for plain EN 16931 CII: no Factur-X profile schema is checked. It carries no verdict and no check depth. A converted document that fails comes back as a tool error that lists the failed rules.

A PDF (Factur-X / ZUGFeRD) must be passed by documentPath for validate, parse, and convert, not inlined as text.

Outages and repeated calls

A tool call retries a short failure of the beliq API for up to 10 seconds, and gives each attempt up to 45 seconds, so a call ends within 55 seconds: under the 60-second request limit that the MCP TypeScript SDK's client sets by default. When beliq asks for a longer wait, as it does with Retry-After: 30 during a database outage, the call does not wait. It returns a tool error that says beliq is unavailable for now and, when beliq sent a Retry-After, in how many seconds to try the same call again. Any other API error that carries a Retry-After, such as a 429 rate limit, names that wait too. When the MCP client cancels a call, the call stops at once, also in the middle of a wait between two attempts.

beliq does not charge a repeat of a charged validate, parse, generate or convert request within 15 minutes of the charge, up to 10 repeats; when it cannot read or write its record of charged requests, it charges every attempt. The 15 minutes and the 10 repeats are beliq's own choice; no standard sets them. Repeated requests are not charged again says what makes two requests the same. The generate and convert tools stay marked as not idempotent, since a repeat after that window is charged again.

Agent skill

skill/SKILL.md is a portable agent skill that teaches a model when to validate, how to read errors/warnings, and how to report a verdict with its check depth, using the tools above. Drop it into a skills directory for an agent that should validate invoices on request.

Development

This server depends on the published @beliq/sdk, which carries the request, transport, and result-shaping logic. package-lock.json is committed, and CI and the release build install from it with npm ci.

  • npm install
  • npm run build - compile to dist/
  • npm run typecheck - src and tests
  • npm run lint - ESLint with type information, whole repository
  • npm test - unit tests (result summary) and an in-memory MCP round-trip with a fake SDK client
  • BELIQ_API_KEY=your-key npm run test:integration - live smoke tests against the real API

Run the built server directly for a quick check:

BELIQ_API_KEY=your-key node dist/index.js

License

MIT

NPM
beliq-mcp
Install Command
npm install beliq-mcp