# TrueForge

An open-source agent harness that runs the full agent execution loop, including model calls, MCP tools, skills, sandboxing, and approvals, behind a chat UI or an API.

TrueForge handles the parts of running an agent well that are easy to skip when only building the agent itself: streaming, session persistence, tool servers, sandboxing, approvals, and a UI, exposed as a bundled chat UI, an HTTP API with a TypeScript SDK, and an embeddable UI SDK. Initial setup configures models, MCP servers, and a sandbox once from shipped, customizable YAML catalogs, after which agents pick from what has been connected; it works with OpenAI, Anthropic, Google Gemini, other catalog providers, or any OpenAI-compatible endpoint.

Its sandbox runs isolated code and file execution through a provider such as Daytona, provisioned only when needed, with secrets kept in the harness rather than passed to the sandbox, and its context engineering features include subagents, deferred tool loading, a code mode, large-result offloading, and compaction. It scales from a single local process backed by SQLite to a hosted deployment with Postgres, Redis, and Docker Compose, Helm, or Railway.

It is written in TypeScript and distributed under the MIT license.

Website: https://trueforge.dev