TrueFoundry, a San Francisco-based enterprise machine learning startup founded in 2021 by former Meta and Google engineers, has published TrueForge, its own agent harness, on GitHub under the permissive MIT License. The harness is vendor-neutral: it can be used with any model a developer or organization prefers, forked, modified, self-hosted and embedded into commercial products. Model, sandbox and infrastructure costs remain separate from the open-source harness itself.
What TrueForge aims to do and why it matters for enterprises
TrueForge is intended to give developers more control over agents and their tools while reducing costs as AI agents become more widely used in enterprises. TrueFoundry reports that when TrueForge was paired with the open source GLM-5.2 large language model, it successfully completed 11 of 14 tasks on DevRev’s Enterprise-Bench (a benchmark testing multi-step tool use across CRM, issue tracking and document management systems). In that test the company reports the cost was 75% lower than achieving the same results with Anthropic’s Claude Managed Agents harness running Claude Opus 4.8: $2.90 versus $11.80. Using the same model in both harnesses (Opus 4.8), TrueFoundry claims roughly a 30% cost reduction for TrueForge compared to Claude Managed Agents ($8.50 vs $11.80).
Anuraag Gutgutia, TrueFoundry’s co-founder and COO, told VentureBeat that many customers had asked for the ability to launch and manage agents themselves. He framed TrueForge not as a replacement for cloud-managed or vendor-managed runtimes, but as a vendor-neutral, lower-cost option organizations can run alongside those services.
How TrueForge fits with TrueFoundry’s product strategy
TrueFoundry’s commercial offering, the AI Gateway, is a control plane that centralizes model and MCP access, credentials, permissions, budgets and observability. TrueForge operates above that layer: it runs the agent loop that lets a model reason, call tools, receive results and continue until a task is finished.
Developers can start locally with a single command and SQLite and then transition the same harness to a shared deployment using Docker Compose or Helm with Postgres and Redis. TrueFoundry warns the local configuration is meant for developer machines and is not intended as an internet-facing production service.
Gutgutia said the company’s long-term aim is for the AI Gateway to be the common layer under whichever agents and harnesses an enterprise chooses. Some organizations will use TrueForge to launch managed agents, others will keep using Claude, other open-source harnesses or internal systems — but TrueFoundry expects traffic to flow through its gateway when governance is required.
Technical approach: context management and cost savings
TrueForge centers on context engineering — controlling how much information is sent back into the model at each step of an agent run. Techniques described by TrueFoundry include:
- Delaying loading of MCP tool schemas until they are actually needed.
- Delegating isolated subtasks to subagents.
- Offloading oversized tool results into files rather than stuffing them into the active context window.
- Processing structured results through code.
- Automatically compacting long-running conversations (context compaction); the documentation sets a default compaction threshold at 50,000 tokens, adjustable per agent.
TrueForge also treats the sandbox differently from runtimes that keep an agent inside an isolated environment for the whole run. The core agent loop stays on the TrueForge server; a sandbox is provisioned as a tool only when the agent needs to execute code or handle files. TrueFoundry says this reduces unnecessary compute and enables a server to run more agents concurrently. The company argues these choices directly reduce model spending.
Comparison with other orchestration harnesses
TrueFoundry compared TrueForge to several other harnesses and managed runtimes. Key points:
- License: TrueForge is MIT; Anthropic Claude Managed Agents is proprietary. Several other harnesses (DeepSeek, LangChain components) are also open source.
- Pricing: the TrueForge harness is free, but model and infra costs are separate. Anthropic charges for tokens plus $0.08 per running session-hour; other solutions’ costs depend on chosen models and infrastructure. OpenAI has estimated developer-level spending around $100–$200 per month in some cases, though actual costs vary.
- Model flexibility: TrueForge supports bring-your-own-model; Claude Managed Agents are Claude-centric.
- Deployment: TrueForge supports a local single-process start and full production deployments with Docker Compose or Helm; Anthropic manages its runtime.
TrueFoundry positions TrueForge’s main differentiator as the combination of an open-source, vendor-neutral harness with a clear progression from local development to shared production runtime, plus an optional enterprise governance layer through TrueFoundry.
Open source versus governance
TrueFoundry emphasizes that using the open-source TrueForge alone does not automatically apply an organization’s enterprise access policies. If an organization runs the open-source harness by itself, it must implement appropriate controls either in the harness or in front of it. When paired with TrueFoundry’s AI Gateway, agents can inherit identities and access controls already attached to models, MCP servers, tools, skills and other agents, allowing central enforcement of SSO, identity providers and fine-grained permissions rather than reimplementing them for every agent.
This distinction is likely important for platform engineering teams evaluating the project: TrueForge is free software, while TrueFoundry’s governance layer is a commercial control plane.
Early users and company background
TrueFoundry said NetApp was a beta user and contributed requirements during development; NetApp’s IT organization used the technology for incident response and faster ticket triage and exposed internal agents as self-service tools for developers. Automattic was also named as an early user.
TrueFoundry was founded in 2021 by Nikunj Bajaj, Abhishek Choudhary and Anuraag Gutgutia. In February 2025 the company raised a $19 million Series A round led by Intel Capital with participation from existing investors Eniac Ventures and Peak XV’s Surge, Jump Capital and angel investors including Gokul Rajaram and Mohit Aron, bringing total outside financing to about $21 million.
The company reported that its customer base had grown fourfold year over year and that by 2025 it was managing more than 1,000 clusters for machine learning workloads. In VentureBeat’s January 2026 coverage of TrueFoundry’s TrueFailover launch, the company said it had more than 30 paid customers worldwide, had exceeded $1.5 million in annual recurring revenue during the prior year, and was processing more than 10 billion requests per month through its AI Gateway.
In June 2026 TrueFoundry acquired UK-based Seldon AI, a longtime MLOps vendor whose Seldon Core software has been used for production model serving and inference. The acquisition reinforces TrueFoundry’s strategy of unifying traditional ML, LLMs, tools and agents behind a common deployment and governance layer.
Conclusion
TrueForge extends TrueFoundry’s product strategy upward into the agent runtime by providing an open-source, vendor-neutral harness that aims to reduce model costs through context management and sandbox-on-demand provisioning. Organizations can run TrueForge independently or alongside other harnesses; TrueFoundry’s commercial opportunity is to provide the governance and control plane beneath whatever agent runtimes enterprises choose to use.
