TL;DR — Key Takeaways
- TrueFoundry says its open source TrueForge agent harness can reduce token costs by managing agent workflows more efficiently than provider-managed harnesses.
- In company benchmarks, TrueForge cut costs by about 30% when used with Claude Opus and by 75% when paired with the open GLM-5.2 model.
- Savings come from reducing unnecessary context, summarizing tool responses, limiting tool calls and handling more orchestration outside the LLM.
Should model harnesses be vendor neutral? They will be if they can save the user money.
AI infrastructure startup TrueFoundry has released TrueForge, an open source agent harness that the company claims can cut token spend over what the internal model harnesses from the frontier labs can offer.
When used with Claude Opus, for instance, TrueForge was able to complete an identical job for 30% lower cost than what Anthropic itself managed with Managed Agents. Moving the workload to an open model, GLM 5.2, cut costs by 75% with no loss of fidelity.
The Harness Commoditizes the Model
Traditionally, the frontier models run their own harnesses, which are invisible behind the user interface. They offer tools that can supervise and orchestrate multiple agents, via such services as Claude Managed Agents. But there is no reason the user can’t run their own harness and communicate with the LLM through HTTP REST APIs.
On its own, a large language model (LLM) is just a stateless text-prediction engine. The harness provides runtime control (and a testing mechanism) for an LLM, giving it instructions on how to operate tools, keeping track of results, and confining the LLM to an approved environment and actions.
“Agents get interesting the moment they stop being islands: sharing tools, calling each other, carrying a skill from one place to the next,” a TrueFoundry blog post argued.
Plus, the company adds that it is less expensive to externalize the agent supervision than to leave it to the provider.
Using 14 business tasks from DevRev’s Enterprise-Bench, the company found that the jobs would cost 10 million tokens (a current market value of about US$11.80) on Claude Agents, but only 3.8 million (US$8.50) on Opus with TrueForge, and only 3.7 million tokens (US$2.90) with TrueForge harnessing GLM-5.2.
The savings come from better management of the agentic loop, the company asserts. Each round trip involves packing up all the contextual information and sending it to the LLM, resulting in the prolific spending of input tokens. TrueForge handles much of the work that would otherwise be done by the LLM, including the planning, file trimming and sub-agent machinery.
Also, TrueForge trims and summarizes the history and tool responses. It also cuts the number of tool calls to just the essentials.
Finally, savings come from the user being able to pick the most efficient model for the job.
Setup and Execution
TrueForge can be set up to run on a local machine, via a Node.js installation, or by container or Kubernetes for cloud and networked environments.
Out of the box, TrueForge supports OpenAI, Anthropic, and Google Gemini. It also supports any models with OpenAI-compatible endpoints such as GLM, Kimi, and MiniMax. When run locally via node, the user can use a copy of the SQLite database. Larger production environments can also use Postgres or Redis.
The TrueForge runtime has three components. A core server runs an agent loop. An HTTP API and TypeScript SDK connect to the outside world and to the model. A chat user interface is also included.
Third-party tools can also be added, and are easily configurable via YAML and Helm. Daytona can provide a disposable sandbox, so that no AI-generated code can interact with the host machine. OpenUI can be used to create dashboards, forms and tables. Users can also wire in search services from Exa, Tavily, Bright Data or others.
For the harness to run in a production setting, TrueFoundry offers an AI Gateway, which can manage permissions, security guardrails and telemetry.
