TrueForge cuts agent costs, but open source is not governance
TrueForge is an open-source AI agent harness that TrueFoundry says completed the same enterprise tasks at 30%-75% lower cost than Anthropic's Claude Managed Agents. The useful distinction is not simply that the runtime is free. TrueForge separates the agent execution loop from the enterprise governance layer TrueFoundry sells around it, while the headline cost comparison comes from a company-reported benchmark rather than an independent evaluation. That makes the project relevant to a real infrastructure decision: whether lower model spending and vendor flexibility justify taking responsibility for the runtime and its controls.
TrueFoundry released TrueForge under the permissive MIT License on Github. The harness can be used with different AI models, forked, modified, self-hosted, and incorporated into commercial products. Its stated purpose is to manage the loop in which a model reasons, calls tools, receives results, and continues until a task is complete. That places TrueForge above the model access and policy functions handled by TrueFoundry's commercial AI Gateway. The separation gives the company a credible open-source entry point without removing the commercial role of its existing platform.
The cost figures come from a specific Enterprise-Bench evaluation conducted by TrueFoundry. Paired with the open source GLM-5.2 LLM, TrueForge successfully completed 11 of 14 tasks covering multi-step tool use across CRM, issue tracking, and document management systems. TrueFoundry says that run cost $2.90, compared with $11.80 for Claude Managed Agents powered by Claude Opus 4.8, a 75% difference for the reported task set. Using Claude Opus 4.8 in both harnesses, the company reports a smaller gap, $8.50 for TrueForge versus $11.80 for Claude Managed Agents, or roughly 30% lower cost. These numbers show that orchestration choices can affect spending, but the source does not describe the full evaluation methodology, so they do not establish a general cost advantage across agent workloads.
TrueForge targets context management as the place where that spending can be reduced. The harness delays loading MCP tool schemas until they are needed, delegates isolated work to subagents, moves oversized tool results into files, processes structured results through code, and compacts long-running conversations. Its documentation sets the default compaction threshold at 50,000 tokens, with per-agent configuration available. Each choice changes how much information returns to the model at each step. That may reduce token consumption when tasks generate large intermediate results or expose many tools, but the source does not isolate the savings from each technique, leaving the benchmark unable to show which design decision carries the largest effect.
The sandbox architecture makes a similar tradeoff. Rather than keeping an agent inside an isolated environment for its entire run, TrueForge keeps the core agent loop on its server and provisions a sandbox as a tool when code execution or file work is required. TrueFoundry argues that this avoids unnecessary compute and allows a server to run more agents concurrently. The claim is plausible within the architecture the company describes, but the source does not provide a concurrency measurement or a comparison under equivalent workloads. The design therefore supports a narrower conclusion: TrueForge makes sandbox use conditional, while the efficiency benefit remains a company claim that still needs workload-specific validation.
The deployment path is part of TrueFoundry's positioning. A developer can run the local configuration with a single command and SQLite, then move the same harness into a shared deployment using Docker Compose or Helm with Postgres and Redis. TrueFoundry explicitly warns that the local setup is intended for a developer's machine and not for an internet-facing production service. That warning is more important than the local-start claim because it prevents the free package from being mistaken for a ready-made public service. The path from local execution to shared deployment exists in the source, but the source does not establish how much engineering work separates those environments.
The comparison with other harnesses reinforces the vendor-neutral pitch rather than proving a universal winner. DeepSeek Harness is described as an MIT-licensed developer preview with replaceable models, tools, storage, and agent loops. OpenAI Codex CLI is open source and focused on software-engineering workflows, while LangChain Deep Agents offers a general-purpose path connected to LangGraph and LangSmith. Anthropic's Claude Managed Agents instead provide a proprietary, Claude-centric runtime managed by Anthropic, with Claude token charges and a $0.08 per running session-hour fee. TrueForge's distinction is the combination of MIT licensing, model flexibility, and a path toward a shared runtime. The comparison does not put every harness through the same benchmark, so it supports a product-positioning argument rather than a category ranking.
Governance is where the free runtime stops being equivalent to the paid platform. TrueFoundry's Anuraag Gutgutia told VentureBeat that an organization running only the open-source harness must provide the appropriate controls through TrueForge or another internal control system. When connected to the AI Gateway, the company says agents can inherit identities and access controls associated with models, MCP servers, tools, skills, and other agents. The gateway also handles enterprise SSO, identity providers, permissions, observability, budgets, guardrails, and failover, according to the source. Those capabilities are not automatically included because the harness has an MIT license. The commercial gateway is therefore not an accessory to the open-source project; it is the layer that turns a flexible runtime into part of a centrally governed platform, if the company's integration claims hold in a customer's environment.
That structure explains why TrueFoundry can give away the execution layer while still pursuing revenue from it. Gutgutia said customers had asked for a way to launch managed agents and MCPs in a vendor-neutral, lower-cost setup, and described TrueForge as complementary to cloud-managed and commercial-provider-managed harnesses. He also said the company's longer-term goal is for AI traffic from different harnesses to pass through its gateway. This suggests that TrueForge is both an engineering response to orchestration costs and a distribution path for the governance plane beneath them. The open license lowers the barrier to adoption, while the gateway creates the commercial control point TrueFoundry wants to standardize around.
TrueFoundry cites NetApp as a beta user that contributed requirements and used the technology for incident response, ticket triage, and internal self-service agents. Gutgutia also identified Automattic as an early user. Those examples indicate that the project has been tested in enterprise-related settings, but the source presents them through TrueFoundry and does not independently measure the reported outcomes. The company's broader history, including its MLOps origins, roughly $21 million in outside financing, more than 30 paid customers reported in January 2026 coverage, and its AI Gateway business, helps explain the strategy but does not validate TrueForge's performance. Customer names and company scale establish commercial context, not representative evidence.
TrueForge presents a defensible architecture for reducing context overhead and avoiding dependence on one model provider. Its strongest evidence is narrower than its enterprise framing: TrueFoundry reports lower costs on 11 of 14 Enterprise-Bench tasks, and the source does not provide enough methodology to determine whether the gap transfers to other tools, models, or task mixes. The deployment decision therefore depends on two separate validations, one for the harness's cost behavior under a team's own workload and another for whether governance remains adequate without the paid AI Gateway.