TrueFoundry open-sources TrueForge to put its gateway beneath more AI agents

Why it matters

Agent costs depend heavily on the runtime surrounding a model. TrueForge gives enterprises a self-hosted option while creating a distribution path for TrueFoundry's paid gateway.

VentureBeat reported on August 19th that Nikunj Bajaj, Abhishek Choudhary and Anuraag Gutgutia, the three founders of TrueFoundry, had released TrueForge, an MIT-licensed runtime for building and operating AI agents with a developer's choice of models, tools and infrastructure.

The founders are giving away a layer that enterprises increasingly need, while positioning TrueFoundry's paid gateway beneath whatever agents they deploy. In the VentureBeat interview, Gutgutia laid out that strategy directly: "all that traffic should still be flowing through our gateway."

Bajaj and Choudhary previously built machine-learning systems at Meta, while Gutgutia spent about seven years in research and portfolio management at WorldQuant. The three met at IIT Kharagpur and founded TrueFoundry in 2021 after seeing how much specialized infrastructure work stood between a trained model and a production deployment. TrueForge applies that original thesis to agents, where the surrounding runtime can determine cost, reliability and security as much as the underlying model.

Giving developers the harness

An AI agent harness controls the repeated loop around a model: send context, receive a proposed action, call a tool, return the result and continue until the task is finished. It also carries the less visible production work, including session persistence, permissions, approvals, sandboxing and context management.

According to TrueFoundry's launch post, TrueForge packages that machinery into a core server, an HTTP API with a TypeScript SDK and a chat interface that can be embedded in another product. Developers can run it locally with an npx command and SQLite. Shared deployments can use Docker Compose or Helm with Postgres and Redis.

The repository documentation warns that the local configuration is intended for use on a developer's machine, not as an internet-facing production service.

TrueForge supports MCP tools, human approval checkpoints, persistent sessions, subagents and generative interfaces. It can defer loading tool definitions, move large tool outputs into files and compact long conversations instead of repeatedly sending the full history to a model. A sandbox is provisioned when an agent needs to execute code or manipulate files, rather than running for the agent's entire session.

That last design is useful, although it is not unique. Anthropic's architecture for Claude Managed Agents also separates the agent harness, session log and sandbox, allowing execution environments to be provisioned when required. TrueForge's clearer distinction is ownership: developers can inspect, modify and self-host the runtime, then connect it to Anthropic, OpenAI, Google or an OpenAI-compatible model endpoint.

The 75% claim comes with an asterisk

TrueFoundry tested TrueForge using DevRev's Enterprise-Bench, which contains 14 tasks across simulated CRM, issue-tracking and document systems. The tasks require agents to retrieve and combine information such as support tickets, engineering issues, sales records and customer commitments.

In benchmark results published August 18th, 2026, TrueFoundry reported that TrueForge paired with the open-weight GLM-5.2 model solved roughly 11 of 14 tasks at an average cost of $2.90 per run. Claude Managed Agents paired with Claude Opus 4.8 solved roughly the same number at $11.80 per run, producing the claimed 75% saving.

The comparison measures two changes at once. TrueForge uses a different harness and a much cheaper model. The result supports TrueFoundry's case for model choice, since Claude Managed Agents is tied to Anthropic's model family. It does not establish that TrueForge's orchestration alone cut the bill by 75%.

TrueFoundry also ran Opus 4.8 through both harnesses. It reported an average cost of $8.50 per run for TrueForge and $11.80 for Claude Managed Agents, a reduction of about 30%. TrueFoundry attributed the difference to smaller prompts, fewer tool calls and context compaction. It reported that TrueForge consumed 3.8 million tokens per run, compared with 10 million for Claude Managed Agents.

These are TrueFoundry's own measurements. The methodology is documented and the underlying Enterprise-Bench tasks are public, but the cost and completion results have not been independently validated.

Infrastructure, sandbox and model charges also remain with the user. The MIT license removes the harness fee; it does not make operating an agent free.

Free software feeds the paid control plane

TrueFoundry already sells an AI Gateway that centralizes model credentials, access rules, spending limits, routing, observability and MCP connections. TrueForge sits above that control plane and runs the agent loop.

Open-sourcing the harness gives developers a low-friction way to adopt TrueFoundry's architecture without committing to its commercial platform. As deployments grow, TrueFoundry can sell the governance layer needed to manage credentials, policies and costs across those agents. The code is the distribution channel; the gateway remains the product TrueFoundry expects enterprises to pay for.

Gutgutia told VentureBeat that customers had repeatedly asked for a vendor-neutral way to launch managed agents. He also said TrueForge was intended to run alongside commercial and internally built harnesses. That approach makes room for heterogeneous enterprise deployments while keeping TrueFoundry in the traffic path underneath them.

TrueFoundry has raised about $21 million. Its $19 million Series A was led by Intel Capital in February 2025, with participation from Peak XV Partners, Eniac Ventures and Jump Capital. TrueFoundry has since reported more than 30 paid customers, while also reporting annual recurring revenue above $1.5 million and fourfold year-over-year customer growth.

TrueForge extends the founders' pitch from deploying models to governing agents. Its adoption will depend on whether engineering teams trust a young open-source project with a sensitive production layer and whether the reported token savings survive outside TrueFoundry's benchmark. The strategic logic is already clear: Bajaj, Choudhary and Gutgutia are willing to give developers the agent runtime if doing so makes TrueFoundry's gateway the common layer below it.