Eclipse Theia clarifies distinction between Platform, AI harness, and IDE

The Eclipse Foundation has released a detailed breakdown of the Eclipse Theia ecosystem, distinguishing between its three core components: the Theia Platform, Theia AI, and the Theia IDE. The Theia Platform serves as a vendor-neutral foundation for building browser-based or desktop IDEs that can run VS Code extensions.

Building upon this, Theia AI acts as an open-source AI harness, providing the necessary infrastructure—such as context management, tool integration, and model abstraction—to build AI-native applications. Finally, the Theia IDE is a complete, ready-to-use product that implements these technologies, allowing developers to use their own LLM API keys or local models without vendor lock-in or telemetry.

Architecture of the Theia AI ecosystem

• Theia AI Harness: Provides a provider abstraction layer for LLM integration, supporting streaming, tool calling, structured output, and model aliasing (e.g., 'fast' or 'default').
• Agent Framework: Supports custom agents defined via code or workspace prompt files (.agents), allowing for delegation between specialized agents like Architect, Orchestrator, and App Tester.
• Standardized Integration: Implements emerging agent standards including Model Context Protocol (MCP) servers/clients, Agent Skills, and Agent Plugins via the Eclipse Foundation AI Registry.
• Security and Transparency: Features a 'Workspace Trust' model for AI, visible and editable prompts, and the ability to run autonomous agents within the Eclipse Enclave sandbox.
• Extensibility: Enables developers to build domain-specific tools by integrating custom tool functions that call internal APIs and using custom response renderers in the UI.

Why it matters

This modular approach allows software engineers to move beyond general-purpose AI coding assistants by building highly specialized, domain-specific IDEs. By providing an open-source alternative to closed harnesses like Cursor, Theia offers developers greater control over data privacy, model selection, and the underlying agent logic.