# claude-ai-mcp: Anthropic's Hub for Claude.ai MCP Integration Issues > claude-ai-mcp is Anthropic's GitHub repository for reporting bugs, requesting features, and receiving announcements about MCP (Model Context Protocol) integration in Claude.ai. It is a communication and tracking hub, not a software library or an MCP server. **anthropics/claude-ai-mcp** — Report issues related to MCP integration with Claude here. - Repository: https://github.com/anthropics/claude-ai-mcp - Stars: 490 · Forks: 83 - Language: Unknown - License: NOASSERTION - Published: 2026-09-10 · Updated: 2026-09-10 · Language: en - Canonical page: https://hysenlabs.com/projects/anthropics-claude-ai-mcp ## What This Repository Is for and What It Excludes The Model Context Protocol (MCP) allows external servers to expose tools, resources, and prompts that Claude.ai can discover and invoke. When that integration breaks or behaves unexpectedly, knowing where to report the problem matters. claude-ai-mcp is Anthropic's designated GitHub repository for exactly that scope: bugs and feature requests related to connecting MCP servers to Claude.ai, session management behavior, protocol transport changes, tool and resource handling, and OAuth authentication. The README defines the boundaries explicitly. This repository is not the right place for the MCP specification itself: that belongs at modelcontextprotocol/specification on GitHub. MCP SDKs are maintained at the modelcontextprotocol organization. General Claude.ai issues that have nothing to do with MCP go to support.anthropic.com. This scoping matters in practice. An issue filed in the wrong repository is either ignored or redirected, costing time. If you have a problem with how the MCP specification defines a message type, that is a spec question. If you have a problem with how Claude.ai interprets or implements that message type, that is an integration question for this repository. ## MCP Transport Protocols Covered by This Repository Claude.ai's MCP integration supports two transport protocols: SSE (Server-Sent Events) and Streamable HTTP. The README mentions SSE deprecation and Streamable HTTP updates as examples of the announcements posted in Discussions. This tells you that the transport layer is an active area of change, and that switching from SSE to Streamable HTTP is something Anthropic signals through this repository before completing the transition. SSE is a unidirectional server-push protocol where the server streams events to the client over a long-lived HTTP connection. Streamable HTTP is a different transport design suited to the back-and-forth structure of MCP tool calls. The README notes protocol changes, session behavior changes, new MCP capabilities, breaking changes and deprecations, and security updates as the types of announcements to expect. For anyone who has deployed an MCP server and connected it to Claude.ai, these announcements are the earliest signal that something in the integration behavior will change. Subscribing to the Announcements category in Discussions is the lightweight way to stay informed without watching the entire repository. ## Session Management and Connection Lifecycle MCP connections between Claude.ai and an external server are not persistent across all contexts. The README lists session management as one of the covered areas, specifically: initialization, reconnection, and lifecycle behavior. These are the behaviors that determine whether a connection survives a page reload, a network interruption, or a long idle period. Initialization is the handshake that establishes the connection and negotiates capabilities. Reconnection covers what happens when that connection drops: whether Claude.ai attempts to re-establish the session automatically and under what conditions. Lifecycle behavior covers when Claude.ai closes a session intentionally, such as when the user navigates away or the server is removed from the configuration. Problems in this area tend to manifest as an MCP server that appears to connect correctly but then silently stops receiving tool calls, or a server that must be removed and re-added to recover from a failed state. The README's inclusion of session management as a covered topic confirms that these are known complexity points in the integration. ## OAuth, Tool Handling, and What Counts as an Integration Issue Two other areas the README lists are OAuth and authentication, and tool and resource handling. OAuth token handling covers how Claude.ai acquires, refreshes, and presents credentials to MCP servers that require authentication. An MCP server that exposes a paid API or a service with per-user access controls will need OAuth, and the behavior of that flow within Claude.ai is in scope for this repository. Tool and resource handling covers how Claude.ai discovers the tools an MCP server exposes, how it parses tool schemas, and how it invokes tools and processes their responses. A tool that appears in the MCP server's manifest but does not appear in Claude.ai's tool list is an integration issue. A tool invocation that produces a malformed response and causes Claude.ai to display an error rather than using the result is also in scope. The boundary between an integration bug (file here) and a behavior question (check the spec) is not always sharp. A useful rule: if the MCP server works correctly with a different MCP client but breaks with Claude.ai specifically, the issue belongs here. ## How to Report a Bug or Request a Feature The README describes two entry points. For announcements, subscribe to the Discussions tab and select the Announcements category. This is a low-noise channel: Anthropic posts to it for protocol changes, session behavior changes, new capabilities, breaking changes, and security updates. It is not a general discussion forum. For bugs and feature requests, use the Issues tab and choose the appropriate template from the list. The README does not describe the exact templates, but the common expectation for a useful bug report is a reproduction case, the observed behavior, and the expected behavior. The repository includes a SECURITY.md file, which provides the process for reporting security vulnerabilities separately from regular bugs. Security disclosures should follow that process rather than going into a public Issue. ## Where to Find the MCP Specification and SDKs This repository is not the source for building MCP servers. The README gives two external resources: the MCP Specification at spec.modelcontextprotocol.io, which defines the protocol messages, transport rules, and capability negotiation; and the MCP Directory at claudemcp.com, which lists available MCP servers. The modelcontextprotocol GitHub organization hosts the official specification repository and the SDKs. If you are building an MCP server from scratch and need to understand the protocol contract, start with the specification. If you have already built a server and are encountering problems specifically with Claude.ai's behavior as a client, return to this repository. The practical difference between the two is that the specification defines what must be true of all MCP clients and servers, while this repository tracks what is true of Claude.ai's implementation specifically, including deviations, limitations, and planned changes. ## Maintenance Status and License Terms The last push to the repository was on 2026-06-08. The repository is not archived. The top-level directory contains a .github/ folder for issue and discussion templates, a LICENSE file, a README.md, a SECURITY.md, and a drafts/ folder. There are no code files or dependencies. The license is not a standard open-source license. The README states that use is subject to Anthropic's Commercial Terms of Service at anthropic.com/legal/commercial-terms. This is a terms-of-service agreement for Anthropic's products, not a software license permitting redistribution or modification. ## Conclusion Use this repository if you are building an MCP server intended to connect to Claude.ai and run into integration problems, session failures, or authentication issues. Subscribe to Discussions for low-noise announcements about breaking protocol changes. Use Issues to report reproducible bugs with steps and expected behavior. Do not file issues here for the MCP specification, MCP SDKs, or general Claude.ai problems unrelated to MCP. Use the project is subject to Anthropic's Commercial Terms of Service, not an open-source license. ## FAQ ### What is MCP in Claude? MCP (Model Context Protocol) is a protocol that lets external servers expose tools, resources, and prompts to Claude.ai. Claude.ai can then discover and invoke those tools. The claude-ai-mcp repository is where Anthropic tracks bugs and protocol changes for that integration. ### Does the Claude desktop have MCP? This repository specifically covers MCP integration in Claude.ai (the web application). The README does not address Claude Desktop separately. For Claude Desktop MCP support, check Anthropic's documentation. ### What is claude ai mcp? The claude-ai-mcp GitHub repository is Anthropic's communication hub for MCP integration in Claude.ai. It covers session management, transport protocols (SSE and Streamable HTTP), tool and resource handling, and OAuth. It is not the MCP specification itself. ## Sources - [anthropics/claude-ai-mcp on GitHub](https://github.com/anthropics/claude-ai-mcp) - [Issues](https://github.com/anthropics/claude-ai-mcp/issues) - [README](https://github.com/anthropics/claude-ai-mcp/blob/main/README.md) --- Hysen Labs editorial analysis, written from the project's own repository and release notes. Cite the canonical page: https://hysenlabs.com/projects/anthropics-claude-ai-mcp