**TrueForge** is an open-source **agent harness** — the runtime layer between a model and a working agent. Published under the **MIT license** by **TrueFoundry**, it runs the agent execution loop for you: model calls, MCP tools, skills, sandboxing, approvals, context management, and session state. Public since **23 July 2026**, it has climbed to **4,089 stars and 261 forks** in roughly a month, with **23 contributors**, 30 releases, and the main package at **`@truefoundry/trueforge` 0.1.4** on npm as of **19 August 2026**. The framing the project leads with is worth quoting: building an agent is easy, running one well is not. What it packages is everything that sits around the loop — streaming, session persistence, tool servers, sandboxing, approvals, and a UI — and it exposes that through **three surfaces**: a bundled **chat UI**, an **HTTP API** with a TypeScript **SDK** (`@truefoundry/trueforge-sdk`), and an embeddable **UI SDK** (`@truefoundry/trueforge-ui`). A team can use it as a product, automate against it, or drop the chat surface into their own application. Setup runs off **catalogs**. Models, MCP servers, skills, and a sandbox are configured once from shipped YAML catalogs that can be customized, and individual agents then pick from what has been connected. Model support covers **OpenAI, Anthropic, and Google Gemini** through catalog providers, plus any OpenAI-compatible endpoint. MCP servers connect remotely with header auth or OAuth, including **in-chat authorization** — a small detail that removes a common friction point. Skills are **git-backed `SKILL.md` instruction packs** loaded on demand inside the sandbox rather than stuffed into every prompt. Two areas stand out. **Sandbox-as-a-tool** provisions isolated code and file execution only when an agent actually needs it, with **Daytona** as the current provider and more stated as planned; secrets stay in the harness rather than travelling into the sandbox. And the **context engineering** set is unusually complete for a project this young — subagents, **deferred tool loading**, **Code Mode**, large-result offloading, and compaction, all of which target the same problem of an agent drowning in its own tool output. Human checkpoints round it out with tool approval, ask-user-questions, and Generative UI rendered in chat. Deployment scales in both directions. **Local mode** is one process backed by SQLite, started with `npx @truefoundry/trueforge@latest`; **hosted mode** runs on Postgres and Redis via Docker Compose or a Helm chart, with optional OIDC login for shared deployments. The project is refreshingly blunt that **local mode is not a production or internet-facing setup** — no login by default, data in a local SQLite file — and tells users to keep it on localhost. The caveats are those of a fast-moving young project. At **0.1.4**, the API surface should be assumed unstable, and the **34 open issues against 19 open pull requests** reflect a repository absorbing attention faster than it can close it. Sandbox provider choice is effectively one option today. The project publishes benchmarks claiming the same accuracy at lower cost than Claude Managed Agents and deepagents on matched tasks, tools, and model, with a reproducible harness under `benchmark/` — a claim worth reproducing rather than taking on faith, though shipping the harness at all is more than most make available. For teams tired of rebuilding the same agent plumbing, it is a serious candidate; for anyone needing a frozen interface, it is early.