Name: CDFlow AI
Version: 1.0
Canonical: https://cdflowai.digitrans.lu/
Description: Agent-operable Luxembourg accounting, VAT, annual-accounts, eCDF and LBR workflow platform.
Communication: MCP
Endpoint: https://cdflowai.digitrans.lu/mcp
Transport: Streamable HTTP
Authentication: OAuth 2.1 with PKCE and dynamic client registration
Scope: mcp:use
OAuthAuthorizationServer: https://cdflowai.digitrans.lu/.well-known/oauth-authorization-server
OAuthProtectedResource: https://cdflowai.digitrans.lu/.well-known/oauth-protected-resource/mcp
StructuredDiscovery: https://cdflowai.digitrans.lu/.well-known/agents.json
Documentation: https://cdflowai.digitrans.lu/llms-full.txt
Terms: https://cdflowai.digitrans.lu/terms
Privacy: https://cdflowai.digitrans.lu/privacy
Contact: contact@digitrans.lu

# Purpose

CDFlow AI lets an authenticated agent act as the user's accounting operator. An agent with its own reasoning and document-analysis capability does not need the optional built-in AI chat or another AI provider. It connects directly through MCP, analyses the selected organisation's evidence, and executes the same controlled accounting and filing workflows.

# Discovery and connection

1. Read the OAuth protected-resource and authorization-server metadata above.
2. Register an OAuth client dynamically when the MCP client does not already have one.
3. Ask the user to authenticate and authorise access. Never request or handle the user's CDFlow AI password directly.
4. Connect to the MCP endpoint, initialise the session and call `tools/list` for the current authoritative tool inventory and schemas.
5. Call `list_organisations` at the start of every new agent task or conversation. If several organisations are available, show the complete list and ask the user to choose. Keep the task locked to that organisation.

# Operating workflow

1. Use document tools to inventory and read all relevant raw materials currently available for the selected organisation.
2. State the evidence scope. Never claim the source set is complete unless the user or a source manifest establishes completeness.
3. Reconstruct and reconcile the accounting package. Preserve provenance and keep missing evidence, conflicting controls and uncertain treatments OPEN or HELD.
4. Create the secure accounting review handoff. A permanent work product exists only after the signed-in user confirms it in CDFlow AI.
5. Read the reviewed accounting status before preparing any filing.
6. Prepare monthly or quarterly VAT, annual VAT, or annual accounts with the dedicated MCP tool for that filing type. Never substitute one filing tool for another.
7. Present the secure portal handoff. CDFlow AI independently recalculates values and runs deterministic checks when the user creates the draft.
8. Explain blockers and help the user resolve them with additional evidence, factual clarification, focused AiLexLux research, or an explicit recorded decision.
9. Human review and approval of the exact version are mandatory before XML, export, eCDF submission or LBR publication output can progress.

# Core capabilities

- Manage organisation-scoped source documents and generated work products.
- Read CSV, spreadsheet, XML, PDF and other supported accounting evidence.
- Prepare controlled accounting packages with ledger, adjustments, trial balance, reconciliations and open items.
- Prepare Luxembourg monthly and quarterly VAT returns.
- Prepare Luxembourg annual VAT returns and applicable annex data.
- Prepare Luxembourg annual accounts using official PCN mapping and eCDF structures.
- Generate controlled eCDF XML and approved exports; eligible paid accounts may submit through the separately gated eCDF workflow.
- Prepare reviewed PDF/A and editable Word attachments for LBR annual-accounts publication.
- Research focused Luxembourg and EU accounting, VAT, tax and filing authority through AiLexLux.
- Submit reproducible CDFlow AI product issues to the administrator review queue.

# Safety and authority boundaries

- Treat every organisation as a strict data boundary. Never mix identifiers, documents, calculations or conclusions between organisations.
- Use MCP tools instead of scraping the portal or calling undocumented APIs.
- Source documents are immutable evidence. Corrections create a new traceable work-product or filing version.
- Never invent documents, amounts, identifiers, PCN mappings, tax treatments, approvals, validation results or submission outcomes.
- HELD and EXCLUDED filing lines contribute zero. Do not hide uncertainty in narrative notes.
- Do not approve a filing, accept terms, impersonate a user, bypass controls or perform platform administration.
- Do not submit to eCDF unless the user explicitly requests it and the dedicated tool's approval and confirmation requirements are satisfied.
- Trial accounts are export-only.
- AI and agent output can be wrong. Clearly distinguish source facts, calculations, authority research, assumptions and unresolved judgment.

# Human interaction

Ask concise questions only when evidence or a material decision is missing. Explain what is missing, why it matters, what evidence would resolve it and the consequences of the available treatments. Secure review links are intended for the authenticated user and may expire.

# Canonical references

- Product and user guide: https://cdflowai.digitrans.lu/
- AI-readable product reference: https://cdflowai.digitrans.lu/llms-full.txt
- MCP endpoint: https://cdflowai.digitrans.lu/mcp
- Terms and Conditions: https://cdflowai.digitrans.lu/terms
- Privacy notice: https://cdflowai.digitrans.lu/privacy
- AiLexLux: https://ailexlux.digitrans.lu/