TrueForge Open-Sources Agent Execution Framework, Claims 30% to 75% Lower Costs Across 14 Enterprise Tasks

TrueFoundry has open-sourced TrueForge, positioning it as a general-purpose, model-provider-agnostic agent harness. Rather than merely wrapping a single model call, it repeatedly runs a “plan, select tools, execute, return results” loop while handling streaming output, context compaction for long-running tasks, subagents, human approval, and session persistence across reconnections. The project is licensed under the MIT License and can be launched locally with `npx @truefoundry/trueforge`. It also provides Docker Compose, Kubernetes, SDKs, and an embeddable chat UI.

The main architectural distinction is that sandboxes are treated as tools rather than persistent resources provisioned for every conversation. A sandbox is created only when an agent needs to run code or manipulate files, so model-only turns such as querying and planning do not incur the same compute costs. The framework can also connect to external tools through MCP, pausing sensitive operations until approval is granted. Persistent state allows existing tasks to continue after the frontend disconnects or the server restarts. These components can be replaced independently, enabling enterprises to choose their own models, tools, isolated environments, and gateways.

TrueFoundry compared TrueForge with Claude Managed Agents using 14 CRM, ticketing, and document-management tasks from DevRev Enterprise-Bench. According to its published results, TrueForge with GLM-5.2 completed 11 of the 14 tasks at a cost of about $2.90, while Claude Managed Agents with Claude Opus 4.8 achieved the same completion count at a cost of about $11.80. When both systems used Opus 4.8, TrueForge still cost about 30% less. The company attributes the difference to tool filtering, context management, and on-demand resource provisioning rather than model pricing alone.

These results are not sufficient to prove that every workload can achieve proportional savings: the sample contains only 14 tasks, the report was conducted by the vendor itself, and the cross-model comparison changes both the model and the harness. Engineering teams should rerun the evaluation with their own toolsets; separately measure model token usage, sandbox startup, tool latency, and failed-retry costs; and verify that approval state, subagent permissions, and persistent data are genuinely isolated in multi-tenant environments.