TrueFoundry has launched TrueForge, an open-source agent harness that lets developers build and run AI agents using models from different providers, positioning it as an alternative to Anthropic’s Claude Managed Agents.
The San Francisco-based enterprise AI infrastructure startup was founded in 2021 by a team that included former Meta engineers. It initially focused on software for deploying machine-learning models before expanding into generative AI infrastructure.
An agent harness is the software layer that manages how an AI agent interacts with the underlying model and external tools. Anthropic’s Claude Managed Agents provide this functionality as a hosted service for long-running agent workloads on the Claude Platform.
TrueForge can run on an enterprise’s own infrastructure and supports OpenAI and Anthropic models as well as more than 20 additional models, according to TrueFoundry. Developers can bring their own Model Context Protocol (MCP) servers and API keys. The company is also offering a hosted version with usage-based pricing.
TrueFoundry says TrueForge can reduce total agent operating costs by 50%, although its published benchmark shows different savings depending on the model and harness being compared.
In a 14-task DevRev Enterprise-Bench test, TrueFoundry said TrueForge and Claude Managed Agents each completed about 11 tasks using Anthropic’s Opus 4.8 model. TrueForge averaged $8.50 per run, compared with $11.80 for Claude Managed Agents, about 30% less.
The difference widened when TrueForge used GLM-5.2. TrueFoundry said that configuration averaged $2.90 per run while completing about the same number of tasks as Claude Managed Agents running Opus 4.8, representing a roughly 75% reduction.
The benchmark was conducted by TrueFoundry, so the results have yet to be independently validated across larger production workloads.
Separating the agent runtime from the model provider can give enterprises more freedom to change models while retaining control of the surrounding agent infrastructure, according to Pareekh Jain, CEO of Pareekh Consulting.
“TrueForge gives enterprises more control and less vendor lock-in,” Jain said. Companies could, for example, route simpler tasks to cheaper or open-source models while reserving more expensive models for workloads that require them, he said. That could let teams switch models without rebuilding the surrounding tool integrations and governance setup.
The separation could be particularly useful for companies in regulated industries, according to Lian Jye Su, chief analyst at Omdia, because it allows them to integrate their own controls for budgets, access, and observability rather than relying entirely on those provided by the model vendor.
The trade-off is that enterprises have to run more of the stack themselves. Su said that includes maintaining the runtime and ensuring the environment meets regulatory requirements.
Whether self-hosting is cheaper will depend heavily on how the agents are used, Jain said. Costs can rise as agents use larger contexts or repeatedly call models and tools, while self-hosting also brings infrastructure and monitoring expenses.
Su said token consumption is likely to account for the largest share of agent operating costs. Self-hosting could therefore become more attractive when companies can use lower-cost open-weight models and already have the infrastructure and engineering expertise to operate agent workloads, he added.
Jain said self-hosting is more likely to pay off for high-volume agents used continuously, where enterprises have greater scope to route workloads across models and make better use of their infrastructure.
For smaller or unpredictable workloads, managed services could remain cheaper because the provider absorbs much of the infrastructure and operational overhead, he said. Agent harnesses could eventually develop into an infrastructure layer of their own.
Su said that is possible, but the technology has not yet reached the level of standardization needed for the harness to become a fully independent, model-agnostic layer. The industry would need greater agreement on how runtimes handle context, model routing and tool use, as well as broader security and safety requirements.
Jain similarly expects agent harnesses to emerge as an infrastructure category, although he does not expect the market to converge in the same way Kubernetes did around containers. Competing vendors are likely to add capabilities beyond the runtime itself as they seek to differentiate their platforms, he said. Standards such as MCP could nevertheless make models and tools more portable between those platforms, Jain added.