truefoundry/trueforge

What it solves

TrueForge addresses the difficulty of moving from a basic LLM prompt to a fully operational AI agent. While creating a prompt is simple, managing the operational requirements—such as streaming, session persistence, tool integration, sandboxed code execution, and user approval workflows—is complex. TrueForge provides a standardized runtime layer (an "agent harness") to handle these infrastructure needs out of the box.

How it works

TrueForge acts as an execution loop that manages the interaction between the LLM and its environment. It connects to model providers (OpenAI, Anthropic, Gemini, etc.) and integrates with Model Context Protocol (MCP) servers for tools and git-backed "skills" for instructions. It utilizes a sandbox (via Daytona) for isolated code and file execution. The system can be deployed in a lightweight local mode using SQLite or a scalable hosted mode using Postgres and Redis. Users can interact with the agent via a built-in chat UI, an HTTP API with a TypeScript SDK, or an embeddable UI SDK.

Who it’s for

It is designed for developers who want to build and deploy production-ready AI agents without manually implementing the underlying runtime infrastructure, session management, and tool-calling loops.

Highlights

• Comprehensive Runtime: Handles model calls, MCP tools, session state, and context management.
• Sandbox Integration: Provides isolated code execution as a tool, keeping secrets secure within the harness.
• Human-in-the-loop: Supports human checkpoints for tool approvals and asking user questions.
• Flexible Deployment: Offers both a zero-config local mode and a scalable hosted mode via Docker Compose or Helm.
• Context Engineering: Includes advanced features like subagents, deferred tool loading, and context compaction.