TrueFoundry open-sources agent harness to rival Claude, OpenAI

TrueFoundry released TrueForge on September 11, an MIT-licensed agent harness built to run on any model or MCP server instead of locking teams into one vendor's stack. The repo was created July 23 and has 5,405 stars, with release candidates still going out on September 12, per TrueFoundry's announcement.

TrueFoundry is pairing the open release with a hosted, pay-per-usage tier, so the open-source version is also a funnel. The company claims TrueForge cuts enterprise agent costs by 50%. That's TrueFoundry's own number. Nobody outside the company has verified it, and the announcement doesn't say against what baseline it's measured.

What's checkable is the positioning. TrueForge drops into the same slot as OpenAI's Agents API and Claude Managed Agents, minus the requirement to run on one company's models. That's a real option for teams that don't want their agent orchestration tied to a single provider's pricing and roadmap.

On September 11, Salesforce also released a harness, but its version is a six-capability architecture bundled with an AI Control Plane, according to Salesforce's post. OpenAI released its own within that same 48-hour span. Three vendors, three different things called a harness, no shared definition.

Anyone evaluating any of these should treat the vendor's cost or efficiency claim as unverified until someone outside the company reproduces it. Read the architecture doc before the press release. Watch whether TrueForge's release-candidate churn settles into a stable 1.0, or whether the vendor-neutral pitch gets undercut by tighter coupling to TrueFoundry's own hosted tier.