A heart for your assistant

Bookkeeping with AI agents. MCP for Accounting Scalable.

Set up a scheduled task for daily bookkeeping in your own agent workflow. Your personal assistant collects documents from your sources, processes document storage, creates entries and links them to bank transactions. You review results, approve their accounting treatment and finalize entries. Tax returns and filing remain your responsibility.

Delegate daily bookkeeping

Your AI posts entries. You have the final say.

Your personal assistant can automatically handle ongoing bookkeeping every day. Set up a scheduled task in your own agent workflow: collect documents, post them and link bank transactions. You review the result, approve its accounting treatment and finalize entries.9

  1. 01You, once

    Set up the daily run.9

    Define when the task runs, which company it covers and which of your own sources and additional rules apply in your agent workflow. Product-specific foundations are already provided. Give your assistant the access it needs.

  2. 02Your AI

    Collect documents.6

    Your assistant searches for documents in approved folders and emails, retrieves invoices from customer portals and uploads them to document storage. It can retrieve existing unposted documents directly there.

  3. 03Your AI

    Post entries and match bank transactions.

    It reads documents, determines account allocation and creates entries with the appropriate document and bank links. It runs the technical posting preview as part of its workflow. It collects open questions for you.

  4. 04You

    Review, approve and finalize.

    See the completed work, clarify open questions and finalize entries reviewed for accounting accuracy. Tax returns and filing remain your task.

Personal receipt sources and rules

A scheduled task for every day.

Add your receipt sources, tasks and working preferences to the existing agent instructions. Your scheduled agent workflow starts the daily runs. Automation can begin with obtaining the invoice and continue through to the entry in Accounting Scalable.9

Accounting Scalable provides the foundations.

Product-specific instructions, bookkeeping rules and knowledge articles help your assistant work with the app and consider relevant statutory requirements.

Add your own tasks, sources, companies and working preferences. Build on the existing foundations to create your personal AI bookkeeper with deliberately authorised tasks and access.

  • Define the company, document sources and time of the daily run.
  • Set up agent access and permissions for the required sources.
  • Agree on follow-up questions, run reports and your regular review.

Every morning, check the new travel expense documents in my folder and software invoices in the approved inbox. Retrieve missing invoices from the agreed customer portals. Upload documents, post clarified business transactions and link the appropriate bank transactions. List missing documents and open questions for my review. Leave entries unfinalized.

Example of a scheduled task in your agent workflow

When daily business becomes year-end work

Coordinate more topics. Automate many closing entries.10

At year end, clarify additional accounting questions – yourself, with your AI and, where needed, your adviser. The year-end assistant in the App guides you through these topics and handles many entries for you, from depreciation to opening the following year.

You and your professional support

Agree on the fundamentals.

Check fixed assets and useful lives, service periods, accruals, deferrals and tax parameters. Clarify open issues and year-end figures. Your AI can prepare figures and instructions for this.

The year-end assistant in the App

Calculate and generate entries.

Automate straight-line depreciation and pooled assets, accruals and deferrals with reversals, tax provisions or receivables, and VAT account balancing. The assistant also generates technical closing and opening entries.

You

Review previews and approve the year-end accounts.

Review previews and trigger each automation in the App. Check the balance sheet and income statement, approve documents and complete your tax returns and filing.

Prepare annual tax returns with your AI.

Corporation tax and trade tax returns are typically annual tasks for each company and assessment year. Your own AI assistant reads the required annual-accounts data through MCP, asks for additional details and can prefill suitable Mein ELSTER forms using separately authorised browser access. You review and approve; actual submission is a separate step.

View the year-end assistant's automation features

An open MCP server for your assistant

Your AI. Your way of working. Your decision.

Accounting Scalable provides the MCP server. You bring your personal assistant and choose its model, tools, scheduling and data processing.

Free choice of model.13

Use an MCP-capable assistant with the model of your choice – as a cloud service or with AI in your own environment. Select new models through your assistant; your bookkeeping stays connected through MCP.

AI access for your daily work.7

Use your existing AI access for bookkeeping, writing, research and other tasks. You do not need an additional AI subscription from Accounting Scalable for the MCP connection.

You decide about your data.8

Choose which AI you trust with documents and bookkeeping data. Local or self-hosted AI can keep model processing in your environment. Or use the provider you already trust.

E-invoice checks in your daily bookkeeping run

Your AI checks the invoice. You decide unresolved cases.

In your configured MCP agent workflow, your assistant considers e-invoice requirements before creating an entry. If suitable structured invoice data is missing, it checks the information available for an applicable transition rule or exception.11

  1. Identify the format

    Consider both ZUGFeRD embedded in a PDF and standalone XML invoices. Missing embedded XML alone does not establish a breach of the requirements.

  2. Assess the requirements

    Consider the supply period, the transaction concerned and known transition rules or exceptions together. An unknown supplier turnover remains unknown.

  3. Ask a focused question

    If an e-invoice is required, or the decisive assessment remains unresolved, the agent holds that receipt for clarification with you. It can continue processing other resolved receipts.

Example clarification request
This invoice is missing suitable structured data. The information available indicates that an e-invoice is required; no applicable exception is documented. Is it available separately, or can you add information about the transaction? The proposed entry remains open until this is clarified.

You remain responsible for the legal assessment. The agent does not replace an unknown value with zero or treat input VAT as lost solely because an invoice is a PDF.

Make time and cost tangible

What two fewer minutes of review could mean.

Assume you process 100 invoices per month. If your active time for format review and clarification falls from three minutes to one minute per invoice on average, the worked example is:

Manual review
5 h 0 min
With agent assistance
1 h 40 min
Illustrative time saving
3 h 20 min
At an hourly time value of €60
€200Time value per month

Assumptions: 100 invoices × (3 − 1) minutes = 200 minutes. 200 ÷ 60 × €60 = €200 of time value per month.

Illustrative model, not a measured customer saving. The one-minute assumption includes the average human review and clarification effort assumed in this example.12

DEMO AND DATA ACCESS

Connect the demo to your AI assistant.

Connect your assistant directly to the public demo. The MCP address and public read token are provided here, with no registration or credential request. Access is limited to reading synthetic sample data.

How to set up read access

Enter the demo MCP address and public read token in your MCP-compatible client. Use Streamable HTTP with Bearer authentication.

The public demo read token is already provided. No personal registration or credential request is required for this access.

  1. Add the demo MCP address as a separate connection in your assistant.
  2. Copy the public demo read token into your client’s Bearer field or a designated environment variable. Do not use personal instance tokens for the demo.
  3. Use Streamable HTTP. If your client requires manual headers, set Authorization to Bearer followed by a space and the public read token. Do not use OAuth sign-in or append the token to the URL.
  4. Check capabilities_get and explicitly select a company from companies_list. The demo agent can read all synthetic companies in this demo.

The demo App MCP is intended for synthetic sample data. For your own data, use the MCP address and actively configured agent of your personal instance. Personal instance tokens belong only in the private client.

Select Bearer authentication and enter the public demo read token. For manual headers, use Authorization: Bearer followed by a space and the token. This public access applies only to the read-only demo.

Request help connecting to the demo MCP →

HTTP 401: check the demo MCP address, public read token and Authorization Bearer header. Contact us for connection help. Personal instance tokens remain private.

A first task for your AI

Explore only approved synthetic data in read-only mode. Check capabilities_get and list companies and their IDs with companies_list. Explicitly select the company I have confirmed. Show three existing entries and a document link. Do not invent IDs or perform any write or send action.

The first read requests in detail

After initialize, discover current schemas with tools/list. companies_list({}) returns the companies. journal_entries_search({company_id: "<bestätigte ID>", limit: 3}) reads existing entries; journal_entry_get({company_id: "<dieselbe ID>", journal_entry_id: "<gefundene ID>"}) reads their details. The actual schemas govern use.

After initialize, discover current schemas with tools/list. companies_list({}) returns the companies. journal_entries_search({company_id: "<bestätigte ID>", limit: 3}) reads existing entries; journal_entry_get({company_id: "<dieselbe ID>", journal_entry_id: "<gefundene ID>"}) reads their details. The actual schemas govern use.

The website MCP at https://accounting-scalable.de/mcp provides product knowledge and needs no token. The demo-app MCP is a separate data space containing synthetic data. Your personal app MCP requires your own explicitly configured agent. None of these data connections alone provides browser control or ELSTER submission.

Website MCP address: https://accounting-scalable.de/mcp. No token required. Select a connection without authentication in your client.

Bookkeeping with AI

Which assistant would you like to connect?

Choose the specific client. The MCP connection provides data and supported actions; browser control and your professional approvals are added separately.

Choose a setup method

Selected: OpenAI Codex CLI

Distinguish ChatGPT, desktop extensions and other agents

Codex CLI is not a ChatGPT web plugin. Claude Web connectors are not local desktop extensions. For each client, check the actual transport, Bearer authentication and separate browser functionality. General MCP advertising does not prove a suitable Accounting Scalable connection.

OpenClaw, Hermes, Grok Bot and Manus each have their own access and tool methods. Use the read/browser permissions actually offered and do not infer compatibility from a product name.

Choose the right MCP connection method

Codex CLI and Claude Web use different setup methods. Select the specific client and first check its requirements for your private read access.

ClientDocumented methodYour next step in the client
OpenAI Codex CLIHTTP MCP with a Bearer token from a private environment variable.

Your own enabled Codex access and private agent read access are required; the following read test must succeed in your client.

Open the new Codex session, check the tool list and read a report with the company and period explicitly confirmed.
Claude Web — remote connectorRemote connector with a Request headers field enabled according to the organisation.

A visible Request headers field and publicly reachable HTTPS app address are required. A cloud connector cannot reach your localhost.

If the Request headers field is missing, this method does not work with the app’s fixed Bearer token. After connecting, check the company, period and values read back.

A registered connection alone does not confirm that the correct data has been retrieved. Compare the app MCP response with your company, period and authorised permissions. The public demo MCP requires separately supplied read access.

Work on forms additionally requires suitable browser access and a clarified professional classification. MCP data access alone does not operate an ELSTER form or submit a return.

Speed in your bookkeeping

Respond promptly. Work with fewer detours.

1. Technical speed: shorter waits

What matters while working is how quickly a view loads and responds to input: opening a document, filtering the posting list or selecting another company. Already loaded views can reuse existing resources; an initial request loads additional data and program components.

2. Faster workflows: fewer detours

Work on the document, entry and bank link in one application. Switch work steps in the same context instead of manually passing files between tools. The company remains explicitly selected; professional reviews and approvals remain your responsibility.

  • Open a document

    Open an existing document from document storage or its linked entry. Check the original and continue to the associated entry.

  • Filter entries

    Narrow the posting list with the available filters and open a result. This lets you check accounts and document links from the same workspace.

  • Switch company

    Select the other company in the company menu. Continue working with its own dataset; switching does not combine separate books.

  • Link a document to an entry

    In an editable entry, select the appropriate existing document and save the link. Check company, transaction and file. Your own changes require a personal instance with write permissions.

Explore the transactions in the demo →

The feature scope

What you can do with it.

Your AI can handle posting work automatically every day through a scheduled task in your agent workflow. You review results and finalize.

Include e-invoice requirements in the review

Your configured MCP agent assesses missing structured invoice data in the context of the supply period, transition rules and exceptions. If a required e-invoice is missing or the case remains unresolved, it asks you and leaves the receipt open.

Have it handled automatically every day

Schedule a daily run for ongoing business transactions in your agent workflow. Your assistant reads bank transactions, documents and existing entries, handles posting work and collects open questions for your review.12

From your sources into document storage

Your assistant uploads PDF, image and XML documents, searches for unposted documents in storage and reads their content. You tell it where to find additional invoices.5

Post documents and link bank transactions

The assistant assigns accounts to documents and creates entries with document and bank links. It runs the technical preview and the retry-safe posting call. Entries remain unfinalized for your review.1

Support invoicing work

Read sales documents and outstanding items, maintain master data and drafts or perform supported document actions through the controlled workflow.3

Support setup and tasks

Have your agent prepare master data and accounts, work on tasks and help you connect banking.

Connect your own assistant

Accounting Scalable provides the MCP server; you choose the assistant and model. Your own revocable tokens determine whether the agent may only read or also create entries.4

Access guides and support

Have your agent find and quote suitable instructions. App support is available in all plans: Basic Support in Solo and Premium Support from Holding upwards.

Support directly in the app.

Create support requests directly in the app and continue the conversation in the integrated portal, without a separate email workflow. Support is available in German and English; scope and priority depend on your plan. Your authorised AI assistant can read cases and replies and, on your instruction with write access, create a request. Continue further communication in the portal.

How to work with it

A clear workflow.

  1. Connect your assistant

    Create your own agent access in your instance with the desired permissions and connect your MCP-capable assistant.

  2. Define schedule, sources and rules

    Set up the daily task in your suitable agent workflow. Define the company, document sources, permitted portals, timing and rules for open questions.

  3. Hand over ongoing bookkeeping

    Your scheduled workflow starts the runs. Then review entries, run reports and open questions. Approve the results for accounting purposes and finalize the reviewed work.

Your questions

Brief answers.

How do website MCP and App MCP differ?

Website MCP provides product knowledge and can request a personal demo on your instruction. With App MCP, your authenticated agent works on your own bookkeeping data.

Which tasks can I give my agent?

Delegate daily document and posting work to it: retrieve invoices, process document storage, post entries and match bank transactions. Then review results and finalize entries.

Can my AI retrieve invoices from a customer portal?

You can instruct your own assistant to retrieve documents from approved customer portals. Its browser tools collect the invoice and transfer it to document storage through MCP.

Can I choose my model and its data processing?

Yes. Choose your MCP-capable assistant and its model – with a provider you trust or AI in your own environment. This lets you decide whom to trust with data passed to the assistant.

Do I need an additional AI subscription from Accounting Scalable?

You do not need an additional AI subscription from Accounting Scalable for the MCP connection. Continue using your own AI access, including for other tasks.

What remains my responsibility?

Review entries, clarify open questions and finalize the reviewed entries. Tax returns and filing remain your task.

Can my AI handle bookkeeping automatically every day?

Yes, through a scheduled task in a suitable agent workflow. Define sources, company, timing and permissions. The workflow starts your assistant, which collects documents, posts entries and creates bank links. You review results and finalize entries.

What happens at year end?

Clarify additional accounting topics and use the year-end assistant in the App for many automatic entries, such as depreciation, accruals, deferrals and opening the following year. Your AI can prepare figures and instructions; operate year-end automation after a preview in the App.

WEBSITE MCP · PRODUCT KNOWLEDGE

Clarify product questions with your AI.14

Your assistant reads the website, explains features and compares plans. It requests a personal demo on your explicit instruction.

Add the address to your assistant's connections. Discuss everything else with your AI.

Personal demos are for business customers. Your details are submitted only on your explicit instruction. You then confirm your email address.

No token required. Select a connection without authentication in your client.

One possible task

Does Accounting Scalable fit my company? Compare the features and plans with my needs. If I want to test it myself, help me request my own demo.

ACCESS & ENVIRONMENT

Your model. Your assistant. Your MCP access.13

Accounting Scalable provides the MCP server. You choose your personal assistant, its AI model and the tools for your document sources – with a cloud provider you trust or in your own environment.

Connect MCP clients individually

  • ChatGPTAI assistant
  • ClaudeAI assistant
  • CursorEditor with AI

Your own model environment – with an additional MCP client

  • OllamaLocal model runtime
  • DeepSeekModel provider

Notes and requirements

  1. The extent of automation depends on business transactions, AI model, available tools and your permissions. Scheduling takes place in a suitable external agent workflow; the website and App MCP are not AI task schedulers. A technical preview is required as a check before creating entries; it does not finalize entries or indicate beta product status. Final professional review, entry finalization and tax returns remain your responsibility.

  2. Agents can access all companies in their instance. Individual agent permissions per company and OAuth 2.1 login are not included.

  3. Use the App for DATEV/audit exports and year-end dialogs. The agent can retrieve report data, information and year-end instructions. Depreciation, accrual, deferral, tax, closing and opening automation are App functions, not approved MCP year-end actions.

  4. Whether a specific ChatGPT, Claude or other client can connect directly depends on the particular client and its supported transport/authentication method.

  5. Documents are limited to 20 MiB per file and your instance's storage quota. You authorize bank access through the bank/provider workflow.

  6. Your chosen assistant provides file, email and browser access through its tools and your permissions. Accounting Scalable handles document storage, entries and bank links through MCP; it does not log into external customer portals independently. The bank link is created when an entry is created. Subsequently matching an existing entry in isolation is currently not a separate MCP call.

  7. MCP is included in all three Accounting Scalable plans. Subscriptions, API calls, models and additional tools from your AI provider are billed under that provider's terms. An existing chat subscription does not automatically cover every API or connector use.

  8. Through MCP, the chosen assistant receives the data retrieved for its task. Its provider, model operation and additional tools determine further processing. Local model processing does not change the hosting of your Accounting instance or external document sources; even a locally started client can use cloud models.

  9. Scheduling and starting a daily run take place in a suitable client or external agent workflow. The website and App MCP server do not provide an AI task scheduler; App tasks and task templates do not start an external AI agent. Automatic runs require continuing valid access, necessary permissions, a reachable instance and tool use permitted by the provider. Frequency, duration, follow-up questions and any required user confirmations depend on the chosen system. A scheduling feature in a chat client alone does not establish MCP access for its scheduled runs.

  10. MCP itself does not operate a browser or manage ELSTER forms. Separate browser access belongs to your assistant. Missing details remain open; review, approval and submission are separate steps. The annual-accounts assistant is an app feature. Its automations run there after preview and your selection; they are not released MCP closing actions. The MCP agent can read report data and closing guidance. Professional closing decisions, approval, finalisation and tax returns remain your responsibility. Depreciation, accruals and tax automations have the procedural limits explained on the annual-accounts page.

  11. The check requires the current e-invoice review workflow in your app instance, suitable MCP tools and a configured agent. It is not individual tax advice or a guarantee of input VAT deduction. Correction, objective evidence and a justified assumption of transitional eligibility may be relevant separately.

  12. The times and hourly rate are illustrative assumptions, not product measurements. Actual effort depends on invoices, clarification, setup and your review. App and AI costs have not been deducted. The time value is not a promised cash saving or tax saving.

  13. The assistant must support the App's MCP transport and authentication. Configure the connection and permissions in the respective client. Plan, administration and client version determine available models, tools and connections. Ollama and DeepSeek require a suitable client environment; a model alone does not provide an MCP client. Logos identify products and belong to their respective trademark owners; they do not indicate partnership or certification. The Blossom mark is owned by OpenAI.

  14. Website access does not read your bookkeeping data and cannot make entries. For later work with your own data, set up the separate agent access for your personal Accounting instance.

  15. The offering is exclusively for businesses within the meaning of section 14 BGB.