AI agents that know how the EU quarter works.

Montoyer is an open-source, domain-specific multi-agent framework that models the internal machinery, civil service workflows, and legislative procedures of the European Union. Unlike general-purpose chatbots, Montoyer breaks down complex institutional processes (Inter-service Consultation, College Deliberation, Trilogue negotiations) into specialised, interacting agents bound by real-world mandates.

No. Montoyer is a completely independent, open-source project built for the Brussels policy ecosystem. All generated outputs are explicitly marked as drafts for official review and do not represent the official stance, legal opinions, or policy positions of the European Commission or any other EU institution.

Honestly? We are doing this for fun, to learn, and to challenge the status quo a little. The EU quarter runs on procedures that are dense, slow, and mostly undocumented for outsiders. We thought it was worth trying to model that, not to replace anyone, but to see what happens when you push on those boundaries with AI agents.

If it helps like-minded people (researchers, builders, curious insiders) think differently about how institutions work, that is more than enough.

We use a strict Inline Attribution Architecture embedded inside native .sh hooks. When an agent generates a text stream, scripts like post_tool_use_citation_matcher.sh and post_tool_use_eurlex_resolver.sh intercept content in real time. They isolate legal citations, cross-reference them against verified local JSON treaty schemas, check live CJEU records via HTTP requests to the Curia server, and inject visible validation or warning tags directly into the text.

That is your own call. If you work at the Commission, the Parliament, the Council, or an agency, you already know the answer to that question better than we do.

What we can say: you can copy and paste any of our prompts into any GPT-like tool you have access to. No guarantees on the output, and the usual draft-for-review caveats apply.

Rather than using a single model to handle everything, Montoyer structures competence into specialised agent families:

• Commissioner Personas: 21 individual portfolio agents (Competition, Trade, Green Deal…) that simulate executive dynamics.
• Role Specialists: Policy officers, legislative drafters, lawyers, and grant managers.
• DG Operational Agents: Profiles modelled after specific Directorates-General to inject technical and operational constraints.

Yes. The /legislative-eu:legislative-proposal command launches a multi-agent orchestration sequence: a policy officer generates a rough brief, routes it through an inter-service consultation with affected DGs and the Legal Service, simulates a College deliberation for political validation, and can model subsequent trilogue negotiation dynamics between mock Parliament and Council agents. Use /simulation-eu:college-deliberation for the deliberation step standalone.

The framework is LLM-agnostic but optimised for advanced developer environments such as Claude Code runtimes. The core of the repository relies on structured file systems, regular expression routers, and local knowledge schemas rather than loose model weights — ensuring the structural framework stays solid regardless of the underlying model used for generation.

Skills are installable capabilities living inside domain plugins (e.g., legislative-eu, competition-eu). They are invoked via namespaced slash commands in your terminal, such as /legislative-eu:treaty-check <proposal> to verify a legal basis or /legislative-eu:impact-assessment <policy brief> to analyse a regulatory path. Each skill is governed by a strict markdown syntax standard.

Compound commands are multi-agent operations that only produce meaningful results because all 21 Commissioner agents are running in parallel with conflicting mandates. A single agent cannot surface structural fault lines. Only a full College can.

• /simulation-eu:mandate-conflict <proposal>: identifies every structurally guaranteed conflict between portfolios, with the legal basis for each position.
• /simulation-eu:red-team-college <proposal>: runs a proposal through all 21 Commissioners and returns only the severe objections. A token-efficient College stress test.
• /simulation-eu:subsidiarity-stress <proposal>: tests the same proposal against five different member-state configurations to find where the subsidiarity check fails.
• /simulation-eu:timeline <proposal>: produces a realistic OLP timeline with blocking dependencies, QMV thresholds, and trilogue risk points.

Yes. The EU's deliberative machinery (subsidiarity tests, inter-service consultation, impact assessments, adversarial College deliberation) is a general-purpose scaffold for high-stakes collective reasoning. It was designed for policy, but it works on any problem where multiple legitimate perspectives must be reconciled under constraint.

Practical non-EU use cases include: corporate strategy reviews where functional leads hold conflicting mandates, standards body proposals that must survive objections from multiple national delegations, research consortium priority-setting, and any decision environment where the quality of the outcome depends on how many well-grounded objections it has survived. The scaffold transfers wherever adversarial rigour matters.

The framework is entirely open to the world. Clone the repository, inspect the file maps, and add new capabilities. Custom skills are defined as structured markdown configurations inside the plugins/*/skills/ directory tree. To contribute, report a bug, or browse open issues, visit github.com/montoyer/eu-agents.