Model Context Protocol (MCP) server for ELSTER (German tax portal) automation via Puppeteer. Submit UStVA, prepare EUER/ESt, sync inbox and history.
Server has 14 well-defined tools with explicit schemas and descriptions. Naming follows verb_noun convention (elster_*_start, elster_*_confirm, elster_session_*). Descriptions are detailed (100-250 chars typical) and explain WHEN to use each tool. However, several parameter descriptions lack specificity: 'report' object in elster_ustva_generate_xml and elster_ustva_start lacks validation rules (e.g., 'amounts must be positive EUR'); 'data' in elster_eur_start and elster_est_start lack field constraints. No output schemas documented, LLMs cannot predict response structure. Error handling is implicit (no recovery guidance in descriptions). Risk annotations (READ_ONLY, WRITE, IRREVERSIBLE) are present but not formalized as toolAnnotations in schema.
Shows the currently loaded ELSTER configuration (with secrets redacted) so you can verify env vars / config.json were picked up.
Starts an ESt 1 A (Einkommensteuererklärung) form-prep session. Opens the form, fills taxpayer basics from config + any extra fields you provide (by ELSTER input id/name hint), runs Prüfung, then waits 30 min for you to review in the portal. NEVER submits.
Starts an EÜR (Anlage Einnahmen-Überschuss-Rechnung) form-prep session. Fills the form up to Prüfung, then tries to "Speichern und Verlassen" so the draft survives in ELSTER. NEVER submits.
Returns the list of supported UStVA Kennziffern (codes 81, 86, 66 etc.) with descriptions and whether they are NET (base amount) or TAX (tax amount).
Verifies that the configured certificate + password can log into the ELSTER portal. Returns success and final URL or an error. Use this once before submitting anything.
Output schemas not documented. LLMs cannot predict response structure for any tool. Downstream tool chaining and field extraction require guesswork.
Parameter 'report' in elster_ustva_generate_xml and elster_ustva_start lacks validation constraints. Description says 'amounts in EUR' but does not specify: must be positive? decimal precision? max value? LLMs will pass invalid values.
Inferred effective spec: <=2025-11-25.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | C | 63 | <=2025-11-25 | v2 |
Cancels a running session (closes the browser, marks status as CANCELLED).
Lists all currently tracked sessions (USTVA / EUR / EST / SYNC) with their status.
Returns the current status, progress log, and any screenshot path for a session started via elster_ustva_start / elster_eur_start / elster_est_start.
Reads "Übermittelte Formulare" (transmission history) from ELSTER. Optionally downloads PDFs.
Reads ELSTER inbox messages ("Posteingang"). Optionally downloads each message as PDF.
Confirms submission of a UStVA session that is in AWAITING_CONFIRM state. Triggers the final "Absenden" click. Returns the transmission ticket on success.
Tests whether a voucher would be detected as Reverse-Charge (§13b UStG) based on the configured supplier patterns. Returns matched supplier and region (EU / NON_EU), or null.
Generates an ELSTER UStVA XML snapshot for archiving. Does NOT submit (submission goes via elster_ustva_start). Useful for audit trails.
Starts a UStVA submission session. Opens a browser, logs in, fills the form, runs Prüfung, then PAUSES at AWAITING_CONFIRM. You must explicitly call elster_ustva_confirm to send. Returns a sessionId — poll status via elster_session_status.
Parameter 'data' in elster_eur_start and elster_est_start is a free-form object with no enum or validation. Description lists 'supported fields' but does not enforce them. LLMs may pass unsupported field names, causing silent failures.
Risk annotations (READ_ONLY, WRITE, IRREVERSIBLE) are metadata but not formalized as toolAnnotations in the MCP schema. Clients cannot programmatically detect destructive tools.
No error recovery guidance in tool descriptions. If elster_ustva_start fails, LLM does not know whether to retry, check config, or call elster_login_test first.