MARVIS-Agent V2 is a local-first workbench designed for comprehensive credit risk model development and workflow management. It targets credit risk professionals, model developers, and data scientists, offering an auditable, governed environment that keeps sensitive work close to local files and runtimes, thereby enhancing security and reproducibility. The agent streamlines tasks from data processing and feature engineering to model validation, deployment, monitoring, and strategy development.

How It Works

MARVIS-Agent V2 operates on a local-first principle, allowing users to serve the platform directly from their machine or server. It features a task-level Agent workbench that guides users through credit-risk workflows via conversational prompts and confirmation gates, maintaining execution context. A key architectural component is its Plugin/Tool/Workflow runtime, which enables the installation and management of modular capability packs with defined schemas, permissions, and auditable outputs. This approach supports both new V2 workflows and the continued use of V1.1 notebook validation runtimes, ensuring backward compatibility and reproducible metrics.

Quick Start & Requirements

• Primary install: Clone the repository, create a Python environment (venv or conda recommended), and install using python -m pip install -e ".[dev]".
• Prerequisites: Python 3.11+ (3.12 recommended), macOS or Linux for verified local workflow, Java runtime (for PMML scoring), Node.js (for frontend syntax checks).
• Local Run: Execute marvis or marvis serve --host 127.0.0.1 --port 8000 --workspace ./workspace and access via http://127.0.0.1:8000/.
• Docs: Core documentation is available within the repository, including plans (docs/plans/v2-master-backlog.md), reviews (docs/reviews/), branding (docs/branding.md), and versioning (docs/versioning.md).

Highlighted Details

• Local-first execution ensures data privacy and control.
• Task-level Agent workbench provides guided, human-in-the-loop workflow execution.
• Plugin/Tool/Workflow runtime facilitates modular and governed capability extensions.
• Supports configurable branding for private deployments while defaulting to a public MARVIS brand.
• Covers a wide range of credit risk workflows: data join, feature analysis, model development, scoring, monitoring, strategy development, portfolio analysis, and vintage analysis.

Maintenance & Community

The project is hosted on GitHub. Specific details regarding active contributors, community channels (like Discord/Slack), or formal roadmaps beyond internal documentation links were not explicitly detailed in the provided README.

Licensing & Compatibility

This project is released under the MIT License, which is permissive and generally suitable for commercial use and integration into closed-source projects.

Limitations & Caveats

The verified local workflow is currently limited to macOS and Linux environments. While V2 represents the active development path for many features, some workflows, like V1.1 model validation, are maintained separately, suggesting potential differences in maturity between components. Updating the agent requires uncommitted local changes to be handled prior to execution.