claude code with jira MCP constant fail!!!

[Dean Hiller]

From claude code side, I have it working with ALL MCP servers I want(sentry, playwright, firestore) EXCEPT jira MCP server which constantly says 

"> /mcp
⎿  Authentication successful, but server reconnection failed. You may need to manually restart Claude Code for the changes to take effect."

I restart and still does not work.

On the flipside, claude.ai works fine but I think uses a different auth process with atlassian perhaps?    so the atlassian MCP server works!!! but only on claude.ai

soooo, is this a jira atlassian MCP server bug for auth or a claude code bug.  linking claude code issue -> https://github.com/anthropics/claude-code/issues/9127

[Clonch, Chris A]

From the linked Claude Code issue (!9127), this is due to the removal of SSE for MCPs in version > 2.0.9 since it was deprecated from the MCP standard (https://github.com/modelcontextprotocol/modelcontextprotocol/pull/206).

Is there plans from Atlassian to upgrade their MCP to use to HTTP or STDIO instead?

I'm much prefer to stay on latest versions of Claude Code _and_ use vendor provided MCPs as much as possible.  I don't like the thought of using some kind of MCP proxy given the risk of supply chain attacks.

[Clonch, Chris A]

Managed to track down the GitHub project and filed a bug -- https://github.com/atlassian/atlassian-mcp-server/issues/8.

[Dylan Wood]

I've seen a lot of reports like this, and ran into it myself one too many times. The issue is that Atlassian's, MCP server uses short-lived OAuth tokens that Claude Code can't refresh, so it works for a bit and then fails — and often gets stuck in a permanently failed state.

I ended up building a proxy that takes a different approach — instead of fighting with OAuth, it lets you use your Atlassian API token to connect via clients that don't support API token auth (Claude.ai, Claude mobile, ChatGPT). It still uses the official Atlassian Rovo MCP under the hood — the proxy just handles auth. 

The key difference is that the proxy handles auth server-side, so Claude Code never needs to manage tokens at all. It also works with Cursor and any other MCP client that supports remote server URLs.

Setup takes about 2 minutes: https://bindify.dev/fix/atlassian-mcp-auth-too-fast  — use code ATLASSIANAUTHFIX for two months free (limited to 10 seats).

Disclosure: I'm the developer behind Bindify.

[Jamie Nelson]

This page shows they know about the problem and shows their workaround.

https://github.com/atlassian/atlassian-mcp-server

TL;DR:

claude mcp add --scope user atlassian -- npx -y mcp-remote@0.1.13 https://mcp.atlassian.com/v1/sse

[Dean Hiller]

SSE is being removed so this is an atlassian bug to port to the newer protocols and is broken until they do!!!

[David Asaf - atd-us]

I run into this every now and then with different clients. Codex. Claude Code. Cursor. 

I usually switch between direct SSE server to proxy like below

"atlassian-mcp": {

"type": "stdio",

"command": "npx",

"args": [

"@playwright/mcp@latest",

"--extension"

],

"env": {}

},

[Dean Hiller]

VERY odd as sentry MCP server uses same protocol and is working so is it claude or jira -> 

"atlassian": {
"type": "http",
"url": "https://mcp.atlassian.com/v1/sse"
},
"sentry": {
"type": "http",
"url": "https://mcp.sentry.dev/mcp"
}