TrueFoundry Gives Away Its Agent Runtime to Sell the Layer Beneath It
TrueForge gives teams a self-hosted way to run the loop around an AI agent. Its commercial bet is that governance, routing and cost controls become more valuable once those agents reach production.
Story brief
3 key pointsTrueFoundry is using an open-source agent runtime to create a paid control-plane business. TrueForge lets teams run modifiable, provider-neutral agents with sessions, approvals, MCP tools and temporary sandboxes, while TrueFoundry’s AI Gateway handles credentials, policy, routing, spending and observability. The company reports lower costs than Claude Managed Agents, including $8.50 versus $11.80 per run with Opus...
- 01
TrueForge is MIT-licensed and deployable locally, with Docker Compose or Helm; local setup is not intended for internet-facing production.
- 02
The runtime supports Anthropic, OpenAI, Google and OpenAI-compatible endpoints, preserving model choice across deployments.
- 03
On Enterprise-Bench, TrueFoundry reported roughly 11 of 14 tasks completed using GLM-5.2 at $2.90 per run versus $11.80 for Claude Managed Agents.
TrueFoundry has released TrueForge, an MIT-licensed runtime for building and operating AI agents. The software lets developers choose their models, tools and hosting setup, while placing TrueFoundry’s paid AI Gateway beneath the agent layer for the controls enterprises may need at scale.
An agent runtime is the machinery that manages an agent’s repeated work cycle: it sends context to a model, receives a proposed action, calls a tool, returns the result and continues until the task ends. It also handles production concerns that do not come from the model itself, including sessions, permissions, approvals, sandboxes and management of a growing conversation history.
The layer around the model
TrueForge packages that work into a core server, HTTP API, TypeScript software-development kit and embeddable chat interface. It supports persistent sessions, human approval checkpoints, subagents and MCP tool support.
How TrueForge tries to limit agent overhead
- It can defer loading tool definitions rather than supplying every tool to every model call.
- It can move large tool outputs into files and compact long conversations instead of repeatedly sending the full history.
- It provisions a sandbox when an agent needs to execute code or manipulate files, rather than keeping one running for the full session.
Teams can run TrueForge locally or deploy it with Docker Compose or Helm, using configurations that support SQLite, Postgres and Redis. The local setup is intended for a developer machine, not an internet-facing production service. The runtime can connect to Anthropic, OpenAI, Google and OpenAI-compatible endpoints.
The business model is not the license
The split is deliberate. TrueFoundry is making the harness available without a license fee, but its gateway remains the product aimed at organizations that need centralized credentials, policy enforcement, routing, spending controls and observability across their agents. The open runtime lowers the commitment required to adopt the architecture; the gateway is positioned to become the shared operating layer as those deployments expand.
That structure also preserves model choice. Claude Managed Agents is tied to Anthropic’s model family, while TrueForge can connect to multiple providers. For a company running agents across several model endpoints, the pitch is not simply a new framework. It is a way to own and modify the runtime while applying a common control layer underneath it.
The cost result with an important asterisk
TrueFoundry tested its system on DevRev’s Enterprise-Bench, a 14-task evaluation spanning simulated CRM, issue-tracking and document systems. It reported that TrueForge paired with GLM-5.2 completed roughly 11 tasks at an average cost of $2.90 per run, versus $11.80 for Claude Managed Agents running Claude Opus 4.8.
That headline comparison changes two variables at once: the harness and the underlying model. It supports TrueFoundry’s argument for selecting a cheaper model, but it does not isolate how much of the claimed saving came from TrueForge’s orchestration.
A more direct company comparison used Opus 4.8 in both systems. TrueFoundry reported $8.50 per TrueForge run against $11.80 for Claude Managed Agents, and 3.8 million tokens per TrueForge run against 10 million. It attributed the difference to smaller prompts, fewer tool calls and context compaction.
The benchmark measurements are TrueFoundry’s and have not been independently validated. The MIT license removes the harness fee, not the costs of models, infrastructure or sandboxes that users still operate.
An adoption test beyond price
TrueFoundry says it has more than 30 paid customers and annual recurring revenue above $1.5 million. TrueForge now extends its earlier infrastructure pitch from deploying models to governing agents, where the runtime has access to tools, files, permissions and long-running task state.
The immediate unresolved question is whether engineering teams will trust a young self-hostable runtime in this sensitive production position—and whether its token and cost advantages persist beyond the company’s benchmark. TrueForge’s release makes the strategic bargain clear: developers get a modifiable agent layer, while TrueFoundry seeks to make its gateway the control point below it.