TrueForge: Open‑Source, Vendor‑Neutral Alternative to Claude Managed Agents

TrueFoundry has announced the open‑source AI agent framework TrueForge, positioned as a direct alternative to Anthropic’s Claude Managed Agents. The company claims that TrueForge can cut total agent operating costs by about 50% and eliminates vendor lock‑in by supporting any model or MCP server.

Vendor‑Neutral Agent Platform

TrueForge enables software engineers to build, deploy, debug, and manage production AI agents on any model—or on MCP servers—without being tied to a single provider. It routes each model call and MCP interaction through TrueFoundry’s AI gateway, allowing budget enforcement, rate‑limiting, and protective measures.

Challenging the Lock‑In Narrative

Co‑founder and former Meta ML lead Nikunj Bajaj argues that hosted‑agent platforms create a misaligned incentive: a vendor selling tokens at $50 per million has no reason to reveal cheaper alternatives that could cost $0.50 per million. This lock‑in forces customers to accept the same tooling, token usage policies, and governance framework as competitors.

Evolution of AI Agents

Bajaj describes a two‑wave evolution: the first wave places agents on developers’ notebooks, integrated into coding tools; the second wave moves agents into customer‑facing products backed by hosted infrastructure. The interface layer becomes a strategic control point that decides when to invoke MCP servers, which models handle which tasks, and how context is retained.

Security and Operational Risks

Some operations still require fully isolated sandboxes, and certain data must never be sent to closed‑source models. Without the underlying framework, development teams would need to implement all of this logic themselves.

Enterprise Governance

When deploying a vendor‑neutral hosted‑agent platform, enterprises must manage persistent sessions, tool credentials, execution sandboxes, context, human approvals, debugging, access policies, and spend. TrueForge’s AI gateway provides centralized control over these aspects.

Illustrative Scenarios

Bajaj uses the placeholder names “foo” and “bar” to illustrate shared‑account pitfalls: a migration from shared to personal access keys caused half the company’s systems to fail because some services still used the old credentials.

Agent Development Lifecycle (ADLC)

TrueForge promotes a unified workload where teams define criteria for models, agents, or MCPs to join a registry. This “Agent Development Lifecycle” enforces consistent operational principles across the organization.

Broad Model and Tool Support

TrueForge supports OpenAI, Anthropic, and over 20 other large language models, plus more than 40 built‑in tools, sandbox execution, human‑review workflows, large‑context handling, generative UI, and Tavily‑powered web search.

Anthropic’s Limitations

According to Bajaj, Claude Managed Agents are limited to a narrow set of Anthropic models, whereas many open models perform better on specific tasks at lower cost. He stresses that Anthropic is not “wrong”; it simply does not offer the full range of models needed for diverse workloads.

Benchmark Results

Using the DevRev Enterprise‑Bench platform’s 14 first‑ and second‑level tasks, TrueForge demonstrated a 50% cost reduction while maintaining comparable accuracy. Savings stem from fewer tokens consumed and the ability to select models outside Anthropic’s catalog that are better suited or cheaper for particular tasks.

Hosted On‑Demand Offering

In addition to the self‑hosted version, TrueFoundry provides a hosted, pay‑as‑you‑go TrueForge service for teams that prefer not to manage the underlying infrastructure.