tramAI: Govern AI Workflows on the JVM

Overview

TramAI is a Kotlin-first runtime for building AI workflows on the JVM. It addresses the gap between making a model call and governing what data, providers, and tools a workflow can use, when a person must approve an action, and what evidence is available afterward.

It is aimed at JVM teams that want typed AI contracts and composable governance capabilities in their application. Standalone and Spring Boot integrations are available; guarantees for routing, approvals, persistence, and evidence depend on configuring the corresponding components.

Key Features

• Define AI operations as typed JVM interfaces, with structured output and validation.
• Apply configurable policy and data-loss-prevention controls before model, tool, and response steps.
• Route requests among configured providers according to workflow classification and policy.
• Gate sensitive work with approval, suspension, denial, and replay-safe continuation capabilities.
• Record audit and execution evidence, with persistence and recovery options when configured.
• Integrate with standalone applications or Spring Boot, and choose provider adapters such as OpenAI-compatible APIs, Anthropic, Azure OpenAI, Bedrock, Gemini, and Ollama.

Use Cases

• JVM teams building an AI-backed business workflow that needs typed inputs and outputs instead of prompt handling scattered through application code.
• Organizations that need policy-based limits on which models or tools can handle classified data.
• Applications where high-risk actions must pause for human approval and resume with execution evidence.
• Teams evaluating local or approved provider routes while keeping AI integration inside a Kotlin or Spring Boot application.

Getting Started

Run the deterministic, credential-free workflow example:

git clone https://github.com/GionaGranchelli/tramAI.git
cd tramAI
./gradlew :examples:governed-workflow:run

See the README for application setup, provider configuration, and further examples.

Alternatives

• mcp-agent: mcp-agent centers on Python agents, MCP server integrations, and composable or durable workflows rather than a Kotlin-first governed runtime.
• shepherd: Shepherd focuses on supervising and replaying agent work as reversible proposals, rather than providing a general typed AI workflow runtime.
• bernstein: Bernstein emphasizes declarative agent policies, isolated task execution, and replayable runs, rather than Kotlin-first typed workflows and provider routing.
• magic: Magic is a self-hostable platform for organizational AI workflows, with collaboration and sandboxing, rather than a Kotlin-first JVM runtime.

Considerations

The project is less than six months old and changing quickly. It has had frequent recent development and releases, but all measured contributor activity is concentrated in one person. The README describes the 0.6.x line as frozen for bug and security fixes, with new features planned for 0.7.0; teams should check the project status and API stability documentation before committing to preview or evolving surfaces.

The sovereign runtime is described as an enterprise proof milestone, not a stable 1.0 API. TramAI does not claim compliance certification or production readiness for every deployment. Governance and evidence guarantees depend on the relevant components being configured. The README also identifies the governed remote MCP tool connector as unimplemented. Building requires a JDK 21 toolchain; the introductory workflow example needs no model credentials, while the OpenAI application example requires an API key.