1/ TrueFoundry, whose AI Gateway sits under production LLM traffic for Fortune 500 companies across healthcare, pharma, financial services and gaming, just open-sourced their agent harness!

TrueForge by TrueFoundry went live on GitHub today. Open Source (MIT License), Vendor-Neutral. That is the permissive kind, with no usage clauses, no commercial restrictions, no community edition with the good parts removed!

You can fork it, run it commercially, and ship it inside your own product.

There are two ways to run it. Local mode runs as a single process with SQLite and no extra infrastructure, intended for personal use on your own machine. Hosted mode runs through Docker Compose or Helm with Postgres and Redis, built for teams and production deployments. The agent features are identical in both modes; only the infrastructure around the server changes.

What you get is an agent harness, the runtime that turns a model into something you can leave running in production.

Tool orchestration, context management, subagents, human approval checkpoints, sandboxed code execution, and turn-by-turn traces of everything the agent did.

TrueForge matched the accuracy of Claude Managed Agents on Opus 4.8, at ~30% lower cost on the same model, and up to ~75% lower when the same work was routed to GLM-5.2.

A thread on what @truefoundry just shipped 🧵: