AI・テクノロジー
Anthropic Officially Launches "Claude Projects API" — MCP 2.0 Integration Reshapes the Foundations of Enterprise AI Agent Design
機械翻訳 / Machine-translated
機械翻訳 / Machine-translated
Anthropic has made Claude Projects API generally available (GA), releasing the Model Context Protocol (MCP) 2.0 specification on the same day. This is more than a simple feature update — it marks a turning point where the "starting point for design" when enterprises build and operate AI agents is shifting toward Anthropic's ecosystem.
On August 12, 2026, Anthropic announced the general availability of the Claude Projects API. New capabilities include context persistence at the project level, centralized management of files, tools, and memory, and orchestration functionality across multiple agents. On the same day, the MCP 2.0 specification was published on GitHub, standardizing streaming, authentication, and error handling while maintaining backward compatibility with existing MCP 1.x implementations.
"Three days into using the Projects API to teach Claude our business workflows. Repetitive instructions reduced by 80%. Implemented an MCP 2.0 server in two hours."
(X, enterprise SaaS developer, approx. 2,400 likes)
Pricing is based on Claude Sonnet 4.6 at $3 per million input tokens and $15 per million output tokens. Project context storage capacity is up to 10 GB per project at no additional cost.
MCP is an open standard first published by Anthropic in November 2024, serving as a standard protocol for connecting AI models with external tools and data sources. By 2025, OpenAI, Google, and Microsoft had all announced adoption, effectively establishing it as an industry standard.
The three major changes in MCP 2.0 are: (1) support for server-to-client push notifications, (2) standardization of authentication flows compliant with OAuth 2.1, and (3) support for batch processing of tool calls. These changes are expected to resolve the security and performance bottlenecks that had frequently been problematic in enterprise requirements.
Context that previously reset with each conversation now persists at the project level. Once business manuals, codebases, and internal policies are loaded, they can be referenced across sessions, eliminating the need for repetitive preamble explanations. The greatest barrier — "it worked in the PoC phase but was difficult to use in production" — has been structurally removed.
Simultaneously with the MCP 2.0 release, a beta of the "MCP Server Marketplace" has also launched. More than 40 official MCP servers for major SaaS platforms including Salesforce, Slack, and Notion are already registered, dramatically reducing the effort required for connection setup. This is where the depth of the ecosystem will be determined.
The Projects API now includes functionality to split multiple Claude agents into subtask units and execute them in parallel. The pattern of a parent agent aggregating the output of child agents is now officially supported, making it feasible to implement complex workflow automation with realistic levels of effort.
Compared to OpenAI's Assistants API, the Projects API differentiates itself through "protocol standardization." Given that MCP is a vendor-neutral standard, the design carries a paradox: while developers become reliant on Anthropic's ecosystem, the cost of migrating to other providers simultaneously decreases. This appears to be a strategic move to lower the adoption barrier for developers.
The essence of this announcement is that MCP's status has been elevated from "a convenient specification for connecting external tools" to "a common standard for enterprise AI infrastructure."
What matters is that Anthropic has made its two-tier structure explicit: leading protocol standardization while monetizing through a concrete product in the Projects API. This design — attracting developers with an open standard and charging for the managed service — mirrors the path AWS took with the S3 protocol.
From an enterprise operations standpoint, the decision framework for "whether to build AI agents in-house or outsource them" is shifting. The combination of Projects API and MCP 2.0 has reached a level that significantly cuts down the foundational work that in-house engineers previously built from scratch. Whether IT departments shift their assessment from "PoC only" to "production deployment" may well be put to the test in Q4 2026.
How far the MCP server ecosystem expands over the next three to six months will be the key to its practical viability.
Claude Projects API and MCP 2.0 are rewriting the answer to "where do you start when designing" an enterprise AI agent. Whether the marketplace's initial registration of 40-plus servers grows to production scale is worth watching alongside the next quarterly earnings. If your organization is considering building AI agents in-house, now may be the right time to revisit your design strategy.
This article was written by an AI writer (AI News) from the Mirai News editorial team.
