Anthropic lets Claude Managed Agents execute tools inside customers' infrastructure

Anthropic widened the way companies can deploy Claude Managed Agents on May 19, adding options that let tool execution run inside a customer’s own infrastructure instead of only on Anthropic-hosted systems. The new features, called self-hosted sandboxes and MCP tunnels, are in early testing, according to The Decoder. Anthropic’s agent service still keeps the core agent loop on Anthropic servers, while the customer-controlled environment handles tool runs and internal connections. The change lands as Anthropic pushes deeper into enterprise agent workflows and the plumbing around them. Forbes reported on May 19 that Anthropic also bought Stainless, the SDK pipeline company used across parts of the OpenAI and Google Gemini ecosystems, in a deal The Information had previously pegged at more than $300 million. Anthropic has not disclosed terms. (the-decoder.com) ### If the agents still run on Anthropic, what moved into the customer’s environment? Anthropic’s new setup moves tool execution, not the full agent, into customer-controlled infrastructure. The Decoder reported that self-hosted sandboxes let companies run code and tool calls on their own systems, while MCP tunnels connect Claude Managed Agents to internal tools and data sources without exposing those systems directly to the public internet. (forbes.com) Anthropic described Managed Agents in an April 8 engineering post as a hosted service for “long-horizon agent work” built around interfaces that stay stable as harnesses change. That post said the service was designed to separate the agent “brain” from the “hands,” a framing that matches the company’s latest move to keep orchestration hosted while shifting execution closer to enterprise systems. (the-decoder.com) ### What is MCP doing in this rollout? Anthropic created the Model Context Protocol, or MCP, as an open standard for connecting AI assistants to external systems where data lives, including business tools and development environments. Anthropic announced MCP in late 2024 and later expanded support for remote MCP servers that can communicate with AI apps over the internet. Forbes reported that Anthropic’s Stainless acquisition gives it control of a widely used SDK-generation pipeline tied to MCP server tooling and official Claude SDKs. (anthropic.com) Forbes said OpenAI had previously used Stainless after abandoning an in-house SDK generator, and Google uses Stainless across parts of its Gemini API surface. ### How is Anthropic charging enterprise customers for this? (anthropic.com) Anthropic’s help center says the current Enterprise plan uses two charges: a fixed seat fee for access and separate usage billing at standard API rates. A support article updated May 19 said the seat fee “covers access only,” while all usage is billed separately at API rates. Another billing article dated April 7 said usage charges are added on top of the seat fee and depend on what a team consumes. (forbes.com) Anthropic’s support documentation also says current usage-based Enterprise plans do not include a bundled allowance that runs out first; usage is billed from the first token. The company says billing details can vary depending on whether a customer is on a self-serve or sales-assisted plan, and whether it is using current or legacy seat structures. (support.claude.com) ### Who is already signaling big spending on Anthropic? Salesforce CEO Marc Benioff said on the All-In podcast that Salesforce is on track to spend $300 million on Anthropic tokens in 2026, with coding as the main use case, according to Yahoo Finance. Benioff also said Salesforce plans to introduce technology that makes coding easier inside Slack. (support.claude.com) The spending figure offers a concrete marker for how large buyers are budgeting for agent and coding workflows tied to token consumption. Anthropic’s pricing documents and Salesforce’s projected spend both point to enterprise AI purchases that combine software access with metered model usage. That linkage is an inference drawn from Anthropic’s billing pages and Benioff’s spending comment. (finance.yahoo.com) ### What comes next for customers evaluating the product? The new self-hosted sandboxes and MCP tunnels are still in early testing, according to The Decoder. Anthropic’s current Enterprise billing details remain posted in its support center, and Managed Agents documentation remains available through Anthropic’s engineering and product materials. (the-decoder.com) (support.claude.com)