Your LLM needs a harness, not another framework

What it does

TrueForge is a self-hostable runtime layer that turns an LLM into a working agent. It manages the full execution loop—streaming model calls, MCP tool servers, sandboxed code execution, human-in-the-loop approvals, and session persistence—and exposes it through a bundled chat UI, an HTTP API, and a TypeScript SDK. You configure models, tools, and sandboxes once via YAML catalogs; agents pick from what you connected.

The interesting bit

Instead of treating sandboxed execution as infrastructure, TrueForge provisions a Daytona container on demand and exposes it as just another tool, keeping secrets inside the harness. It also loads git-backed SKILL.md instruction packs into that sandbox when needed, so an agent can carry specialized context without bloating every prompt.

Key highlights

• Catalog-driven setup: Configure models, MCP servers, skills, and sandboxes in one place; agents consume what you connected.
• Sandbox-as-a-tool: Isolated code and file execution via Daytona, spun up only when needed, with secrets kept out of the agent context.
• Human checkpoints: Built-in tool approval, user questions, and Generative UI inside the chat interface.
• Two deployment modes: Local mode with SQLite for experimentation, or hosted mode with Postgres and Redis for teams.
• Benchmarked: The project publishes reproducible benchmarks comparing cost and accuracy against Claude Managed Agents and deepagents.

Caveats

• Local mode is not production-safe: It runs without authentication and stores data in a local SQLite file; the README explicitly warns against exposing it to the internet.
• Sandbox provider lock-in: Code execution currently relies on Daytona; additional providers are planned but not yet available.

Verdict

TrueForge is worth a look if you are tired of gluing together model APIs, tool servers, and sandboxing just to get a reliable agent loop. If you only need a thin wrapper around a single LLM call, it will feel like overkill.