TrueFoundry formally launched TrueForge on August 19, a day after publishing the project. It is an MIT-licensed agent harness that enterprises run inside infrastructure they already own.
The company aims it at Claude Managed Agents, the hosted agent service Anthropic moved into public beta on April 8. TrueFoundry says the switch cuts total agent operating cost by 50%.
Nikunj Bajaj, co-founder and CEO of TrueFoundry, explained the positioning to me when we spoke ahead of the launch. “We are providing an open source equivalent of Claude Managed Agents,” is how he framed the product.
The harness is the layer most enterprises have thought about least. It is the runtime wrapped around a model that turns one instruction into work, and it plans the task, holds session state, compresses context when the conversation outgrows the window, isolates code execution and stops for human approval before anything irreversible. Managed agent platforms typically bundle their own orchestration runtime, and that runtime rarely travels when the model underneath changes.
Inside The Harness
At its core, TrueForge builds the agent from a plain-English instruction rather than from a file a developer maintains. The user describes the outcome in a sentence or two. The harness asks a model for a plan, then executes it against whatever models, MCP servers, skills and sub-agents the environment has been configured with. It tracks what each tool call returned, decides what to compress before the next model call, and decides when a step is sensitive enough to run inside a fresh sandbox.
Deferred tool loading is the part of that loop worth dwelling on. TrueForge pulls in only the tools the current task needs, so a large catalog does not flood the context window at session start. Bajaj cited MCP servers exposing as many as 2,000 tool definitions as the case that forced the design.
Bajaj does not sell the harness as a reasoning machine that fires a frontier model at every step. “Whatever can be done with deterministic steps, it will want to do with deterministic steps.” When a developer wants tighter control, the graph stays in the developer’s hands and only the reasoning steps are delegated, which should appeal to teams that want deterministic application logic wrapped around bounded reasoning.
TrueForge runs in a single process with SQLite for local work, while shared deployments use Postgres and Redis with either Docker Compose or Helm. Kubernetes is not required and neither are GPUs, since the harness needs only the ability to spin up and destroy isolated containers. GPUs enter only when the customer self-hosts the models.
The governance story becomes stronger when combined with TrueFoundry’s AI Gateway. TrueForge can talk directly to customer-provided models and MCP servers using the customer’s own keys. Enterprises that pair it with TrueFoundry’s commercial AI Gateway can additionally route those interactions through a central layer for budgets, role-based access control, guardrails and unified traces.
TrueForge Against Claude Managed Agents
Anthropic’s product isn’t a thin wrapper, and the comparison is fairer for that reason. Claude Managed Agents pairs a tuned harness with persistent sessions, sandboxed code execution, credential vaults, scoped permissions, MCP support, approvals and execution tracing. Rakuten deployed it across product, sales, marketing and finance. Pricing uses standard model token rates plus $0.08 per active session hour, excluding idle time.
The difference is narrower than it was in April. Managed Agents is an Anthropic-hosted control plane built around Claude, although customers have been able to attach self-hosted sandboxes for tool execution since late May. Anthropic also describes provisioning execution environments only when a task needs them, which is close to what TrueForge does, so the line between the two products is model choice rather than sandboxing. TrueFoundry says TrueForge ships with support for OpenAI, Anthropic, Google Gemini and more than 20 other models alongside more than 40 built-in tools, plus any OpenAI-compatible endpoint.
Bajaj made the ownership argument in the sharpest terms of our conversation. “You’re not dependent on a third party maintaining some of the most important software that is being automatically written in your company,” he said, describing the appeal to enterprises. Self-hosting does not erase that risk; it relocates it, and a platform team now answers for a runtime that writes and executes code.
TrueFoundry says Automatiq and NetApp are running agentic workloads on the harness. Its own Ask TFY assistant runs on the same code, which is a more useful proof point than a customer logo.
Where The Evidence Runs Out
The 50% headline needs qualification, and TrueFoundry deserves credit for making that possible. The company published a reproducible benchmark built on 14 level-one and level-two tasks from DevRev’s Enterprise-Bench, with three trials per configuration and blind model judging. TrueForge came out roughly 30% cheaper than Claude Managed Agents when both ran Opus 4.8, and roughly 75% cheaper when TrueForge ran GLM-5.2 instead. Neither result establishes a 50% cut in an enterprise’s total operating cost once infrastructure and staffing are counted.
The test is also small and vendor-run. GLM-5.2 scored 11.7 out of 14 against 10.7 for Claude Managed Agents on Opus 4.8, an interesting result that warrants independent replication before anyone builds a model strategy on it.
Project governance is the open question. An MIT license on a single-vendor repository is not the same as neutral stewardship, and Bajaj told me foundation membership remains undecided.
Reproducibility is the other thing to press on. TrueForge persists sessions and exposes traces, so this is not a black box. When control flow is generated at runtime rather than written as a graph, though, reproducing a failure six weeks later depends on complete traces, pinned model versions and versioned configuration.
What This Means For Buyers
The first question for enterprise buyers is portability. If the harness is replaced, how much of the agent definition survives, given that there is no code artifact to carry across. The second question is cost attribution, because when tool calls double, someone has to say whether the meter that moves belongs to the model provider or to the infrastructure the customer now operates and staffs.
For the vendors, the pressure is arriving from a direction they did not price for. Open-weight models are increasingly competitive on some agentic workloads. If that holds across harder tasks, the hosted runtime stops being the only credible route into production, and Anthropic, Google and AWS will each have to argue that theirs justifies the infrastructure it ties customers to.
TrueForge is a calculated bet that enterprises would rather operate the agent runtime than rent it. The reasoning is compelling and coherent even where the cost claim outruns the published evidence. It will be interesting to see whether the community builds the vertical specializations Bajaj expects, because that is what would turn a well-designed harness into a platform enterprises across finance, legal and support can standardize on.