TrueForge claims lower Claude Managed Agents cost in two different tests
TrueFoundry released its MIT-licensed agent harness and reported 30% to 75% lower task costs, though the larger comparison also changes models.
By Wei-Lin Zhao · AI Correspondent
· 3 min read
TrueFoundry has released TrueForge, an MIT-licensed open-source agent harness, and says its TrueForge Claude Managed Agents cost comparison showed task-completion costs 30% to 75% below Anthropic’s managed runtime. The larger figure compares different model-and-harness combinations; the company’s same-model result was roughly 30% lower, a narrower claim for teams assessing the runtime itself.
An agent harness runs the repeated cycle in which a model decides what to do, calls tools, processes their results and continues until it completes or stops a task. TrueForge is designed to be modified and self-hosted, while Anthropic operates the runtime for Claude Managed Agents, according to VentureBeat.
How did TrueFoundry calculate its Claude Managed Agents cost claim?
TrueFoundry reported that TrueForge paired with GLM-5.2 completed 11 of 14 tasks on DevRev’s Enterprise-Bench for $2.90. It compared that outcome with $11.80 for Claude Managed Agents running Claude Opus 4.8 and described the result as 75% cheaper. The benchmark covers multi-step tool use across CRM, issue-tracking and document-management systems, VentureBeat reported.
That is a comparison of two stacks, not an isolated measure of harness efficiency: both the harness and underlying model change. For a more controlled comparison, TrueFoundry said it ran Opus 4.8 in both products. It reported a $8.50 cost with TrueForge, versus $11.80 with Claude Managed Agents, or about 30% less.
Neither result is independently validated in the supplied materials. They also do not provide a complete benchmark methodology. A purported TrueForge evaluation-documentation URL returned a 404 page. Operators weighing the claims should treat task success, test construction and cost as linked measures, rather than infer general production performance from a 14-task result. A sound AI model evaluation uses representative work and explicit scoring criteria.
What is TrueForge intended to reduce?
TrueFoundry says its design limits the context sent to a model during an agent run. The techniques include loading Model Context Protocol tool schemas only when needed, passing discrete work to subagents, storing oversized tool output in files, processing structured responses in code and compacting long conversations. VentureBeat reported a default compaction threshold of 50,000 tokens, adjustable by agent.
The company also says a sandbox is created only when an agent needs to execute code or work with files, rather than being held for the full run. Those are product-design assertions; the reported tests do not establish how much savings each mechanism produces.
Open source does not make an agent deployment free
TrueForge has no license charge, but teams still pay for models, sandboxes and infrastructure. TrueFoundry says it can run locally with SQLite, then move to Docker Compose or Helm deployments using Postgres and Redis. It warns that the local setup is for a developer machine, not an internet-facing production service.
Claude Managed Agents is Claude-centric and is listed as charging for Claude tokens plus $0.08 per running session-hour, metered to the millisecond. TrueFoundry positions TrueForge as a vendor-neutral option that can sit alongside commercial and cloud-managed harnesses, rather than as a universal replacement, according to comments from co-founder and COO Anuraag Gutgutia reported by VentureBeat.
This story draws on original reporting from VentureBeat.