Solitarius

An agentic architecture for computer-aided molecular inverse design.

Describe a molecular-design goal in plain language; Solitarius plans the run, configures and executes the generation run, analyses the results against measurable criteria, and decides whether to continue, rerun, or revise, keeping a traceable record from scientific intent to candidate molecules.

---

What is Solitarius?

Generative molecular design imposes a coupled barrier: it needs domain knowledge to translate a scientific aim into a computable objective, and technical expertise to configure a generative workflow that actually reflects that aim. Solitarius addresses both by separating agentic decision-making from molecular-generation execution across three connected layers:

Layer	Role	Repository
1: Agentic orchestration	Coordinates specialist LLM agents through a shared ledger, with explicit approval gates that keep the scientist in the loop.	lcmd-epfl/Solitarius-agentic-system
2: Solitarius MCP	Exposes engine operations as typed, validated tools an LLM agent can call through a Model Context Protocol (MCP) interface.	lcmd-epfl/Solitarius-mcp
3: Engine	REINVENT4 provides molecular generation, transfer learning, and reinforcement-learning optimisation.	MolecularAI/REINVENT4

LLM agents translate user-provided scientific intent into automated generation tasks, scoring objectives, and constraints. A ledger-based orchestration layer records every decision, execution, and outcome, so workflows can be inspected, resumed, and refined.

---

Repositories

→ Solitarius MCP: the execution interface

The MCP server. Exposes the generation engine as LLM-callable tools (validation, scoring-component construction, sampling/transfer-learning/reinforcement-learning execution, job management, monitoring and analysis), each following a dry_run → launch protocol. Works with any MCP-capable platform (Claude Code, Codex CLI, Gemini CLI). Usable standalone or as the execution interface for the orchestration layer.

→ Solitarius-agentic-system: the orchestration layer

The Claude Code-native multi-agent system built on top of Solitarius MCP. Four specialist agents (Planner, Formulator, Executor, Analyser) coordinate through a shared TASK_PLAN.md ledger to run a complete campaign autonomously, pausing only at approval gates.

Install order: Solitarius MCP first, then Solitarius-agentic-system. The orchestration layer uses the MCP server as its controlled execution interface and will not function without it registered and reachable.

---

Campaign coordination

Solitarius treats a design request as a persistent campaign rather than a sequence of independent tool calls. The coding-platform LLM acts as the orchestrator: it reads the current state, calls the next specialist agent, and routes decisions through the shared TASK_PLAN.md ledger.

The four specialists keep the workflow explicit. The Planner turns a natural-language objective into measurable criteria; the Formulator builds or reuses scoring components when the objective needs custom logic; the Executor prepares, validates, and launches generation jobs through Solitarius MCP; and the Analyser evaluates generated molecules and optimisation histories before recommending completion, rerun, or reformulation. Approval gates before plan adoption, launch, and revision keep human scientific judgement in the loop while preserving a traceable record from initial intent to candidate molecules.

---

Demonstrated applications

• Inverse design of singlet-fission chromophores. Four trials (SF1-SF4) preserved one excited-state energetic objective while the structural requirements changed: unconstrained de novo generation (SF1), de novo generation with literature-derived SMARTS exclusion filters applied then expanded (SF2, SF3), and mol2mol lead optimisation from tetracene (SF4). The Executor↔Analyser loop corrected ineffective reward transforms, premature termination, and inverted optimisation gradients without restarting the campaign.
• Targeted dataset expansion: OSCAR!(DHBD) organocatalysts. A "explore underrepresented HOMO-LUMO regions" goal was converted into a transfer-learning → reinforcement-learning workflow with a custom DHBDFMOExplore scoring component, then the same component was generalised to redirect generation across complementary property-space quadrants; no manual workflow reconfiguration between runs.

Generated molecules are surrogate-model candidates and require higher-level (e.g. quantum-chemical) validation.

---

Getting started

1. Install and register Solitarius MCP (local stdio or Docker HTTP).
2. Install the Solitarius-agentic-system orchestration layer.
3. Open Claude Code in the agent directory and state your objective, or run /solitarius <objective>.

---

Citation

If you use Solitarius, please cite:

Solitarius: An Agentic Architecture for Computer-aided Molecular Inverse Design

@article{worakul_solitarius_2026,
    title = {Solitarius: {An} {Agentic} {Architecture} for {Computer}-aided {Molecular} {Inverse} {Design}},
    url = {https://chemrxiv.org/doi/abs/10.26434/chemrxiv.15005834/v2},
    doi = {10.26434/chemrxiv.15005834/v2},
    publisher = {American Chemical Society (ACS)},
    author = {Worakul, Thanapat and Corminboeuf, Clémence},
    year = {2026},
}