TrueForge Releases Open-Source Harness for Running LLMs as Real Agents

TrueForge has released an open-source Runtime Layer that transforms LLMs into fully functional agents. This harness automatically handles model invocation, MCP tool integration, skill management, sandbox execution, approval workflows, context management, and session state. It is exposed via three interfaces: a chat UI, an HTTP API (TypeScript SDK), and an embeddable UI SDK.

Key Features and Architecture

• Integrated Execution Environment: Provides built-in support for streaming, session persistence, tool servers, sandboxing, and approval workflows. Once models, MCP servers, skills, and sandboxes are configured via a YAML catalog, agents automatically select and use them.
• Broad Model and Tool Support: Supports major providers including OpenAI, Anthropic, and Google Gemini, as well as OpenAI-compatible endpoints. It allows on-demand loading of remote MCP servers using header authentication or OAuth, and Git-based SKILL.md skill packs.
• Sandboxing and Security: Offers isolated code/file execution environments using tools like Daytona, with secrets securely stored within the harness. Supports tool approvals, user questions, and human-in-the-loop checkpoints via Generative UI.
• Context Engineering: Efficiently manages context through sub-agents, lazy tool loading, Code Mode, and offloading/compaction of large results.

Deployment and Performance

• Scalability: Supports both local mode (SQLite, single process) and hosted mode (Postgres, Redis, Docker Compose, Helm, Railway), covering use cases from personal testing to team-scale production deployments.
• Benchmarks: Claims to achieve lower costs while maintaining the same accuracy as Claude Managed Agents and deepagents under identical conditions of tasks, tools, and models.