TrueForge opens the agent execution layer as open source

The model still sets the ceiling, but the harness, the runtime layer that turns an LLM into a working agent, is beginning to dictate cost, control, reliability and portability. TrueForge, a new MIT-licensed project, is one of the clearest signs that this category is maturing and moving beyond a collection of ad-hoc scripts.

TrueForge runs the agent execution loop: model calls, MCP tools, skills, sandbox, authentication, context management and session state. It exposes all of this in three ways—a built-in chat UI, an HTTP API with a TypeScript SDK, and a UI SDK for embedding. The idea is simple: building an agent is easy; running it correctly is another story, requiring streaming, session persistence, tool servers, sandboxing, auth and an interface. TrueForge supplies this package out of the box.

Initial configuration is done through YAML catalogs that define models, MCP servers, skills and sandbox once; agents then select from the assembled components. Support spans OpenAI, Anthropic, Google Gemini and any OpenAI-compatible endpoint. MCP tools work with header auth or OAuth, including in-chat authorization. Skills are Git-hosted instruction packages (SKILL.md) loaded on demand into the sandbox; currently Daytona, with additional providers planned. Secrets remain inside the harness itself and are not leaked to the runtime environment.

There are three kinds of human checkpoints: tool approval, user questions, and Generative UI in chat. In the context-engineering side there are sub-agents, deferred tool loading, Code Mode, large-result offloading, and compaction. All of these aim to keep the context window sane without losing critical information. Execution scales from local mode (single process, SQLite) to hosted mode (Postgres + Redis, Docker Compose or Helm) using the same code but different deployment.

The project publishes a benchmark against Claude Managed Agents and deepagents on the same tasks, same tools and same model, same accuracy, lower cost. The reproduction code resides in the benchmark/ directory and the full write-up in Benchmarking. This is a claim that can be tested, not a marketing promise, and should be run independently before relying on it in production.

Developers stress: local mode is intended for your machine only, there is no default login, data resides in a local SQLite file, and you must not expose it to the internet. For shared or production use, only hosted mode is supported. The license is full MIT, the community gathers on Discord, and the founders (Abhishek and Anuraag) provide direct emails for contact. Contributions are accepted per CONTRIBUTING.md, and security reports via SECURITY.md only.