TrueFoundry, a San Francisco-based ML startup founded by former Meta engineers, has open-sourced TrueForge under the MIT License. The agent harness lets developers build, fork, and self-host AI agents that can use any preferred model, aiming to reduce costs and increase control as enterprises scale multi-step AI workflows.
In tests on DevRev’s Enterprise-Bench, TrueForge paired with the open-source GLM-5.2 LLM completed 11 of 14 tasks for $2.90—75% cheaper than Anthropic’s Claude Managed Agents using Opus 4.8 ($11.80). Even with the same model, TrueForge claimed roughly 30% savings ($8.50 vs $11.80). The company positions TrueForge as a vendor-neutral alternative to managed services, complementing its commercial AI Gateway for enterprise governance.
TrueForge’s architecture focuses on context engineering to minimize waste, including delayed tool schema loading, subagents, and automatic conversation compaction at a default 50,000-token threshold. It runs locally with SQLite before scaling to Docker Compose or Helm deployments with Postgres and Redis, though the company warns against internet-facing use in local mode. Early users include NetApp and Automattic, leveraging the harness for incident response and internal tooling.
