Technical overview

Simon is a desktop application that runs a complete accounting engine on your machine. Here is how the pieces fit together.

The 4 layers

flowchart TB
    subgraph Desktop["Desktop application"]
        UI["Graphical interface<br/>Electron + SolidJS"]
        Agent["AI agent"]
    end

    subgraph MCP["Business logic server"]
        Tools["Specialized MCP tools<br/>business modules"]
    end

    subgraph Domain["Business logic"]
        Rules["Accounting and tax rules,<br/>qualification and bank<br/>reconciliation algorithms"]
    end

    subgraph DB["Database"]
        PG[("Embedded PostgreSQL<br/>Local database")]
    end

    UI --> Tools
    Agent --> Tools
    Tools --> Rules
    Rules --> PG

The diagram reads on four levels:

Layer	Role
Desktop application	Electron/SolidJS graphical interface and conversational AI agent.
MCP tools	Specialized business actions: documents, bank reconciliation, qualification, manual entries, returns.
Business logic	PCG rules, tax rules, qualification, bank reconciliation and controls.
Local database	Embedded PostgreSQL, files, views and workflow traces.

When you click a button or when the agent runs an action, both take the same business path.

The database and the files stay local. If you use an external AI provider, or services such as INSEE, VIES or the ECB, the context needed leaves toward the chosen service.

A core principle

The graphical interface and the agent use the same business tools. What the agent can do, you can generally do from the interface, and vice versa. A single source of logic strongly reduces divergences.

---

The agent engine

Simon is built on a fork of OpenCode, an open-source AI agent framework. This choice gives access to:

• Configurable AI providers — Anthropic, OpenAI, Google, compatible providers, local models depending on the available configuration. You choose the model and the provider.
• Declared agents — a main agent orchestrates the work and delegates to three specialized sub-agents when the task requires it: document annotation, document reclassification, and documentary research. FEC import and fixed assets, however, are not separate agents: the main agent drives them through dedicated skills.
• Guided skills — detailed instructions teach the agent how to carry out each accounting task: annotate an invoice, qualify an expense, reconcile a statement, prepare a year-end closing, generate a VAT return…
• Smart session management — the agent keeps track of the conversation and automatically compacts the history when it grows too long, without losing essential context.

Each agent has access only to the tools it needs. The document annotation agent cannot modify your journal entries. The documentary research agent can only read the knowledge base. This isolation limits errors and reinforces reliability.

---

The integrated services

• Company search — querying the INSEE API to find a company by SIREN or by name.
• European VAT validation — checking intra-Community VAT numbers via the European Commission’s VIES service.
• Exchange rates — historical and current rates from the European Central Bank, with an offline mode.
• PDF generation — invoices, manual entries, opening balances, import manifests, directly from the application.
• Tax knowledge base — about 2,000 reference fact sheets drawn from the Bulletin Officiel des Finances Publiques (BOFiP) and the official tax notices. The agent relies on them to answer your regulatory questions and cite its sources.

---

Guiding principles

• MCP symmetry — a single business execution path for the agent and the interface. No hidden logic, no functional shortcut: validations are concentrated in one place.

• Accounting by annotation — most journal entries start from the document: the agent annotates it, qualifies it, and the entries are generated automatically once the data is sufficient. Manual entries remain possible for cases outside the documentary flow.

• Block-based workflow — each document automatically progresses through the stages (extraction, qualification, bank reconciliation, posting). When a problem threatens data integrity, progression stops with an explicit block and a clear explanation of what needs to be fixed.

• Pattern learning — when you correct the qualification of an expense, Simon remembers your choice. The next invoice from the same supplier, in the same category, will be qualified automatically.