[BUG] ### Title Plugin MCP server tools intermittently never register in-session despite `claude mcp list` showing "Connected"

Preflight Checklist

• [x] I have searched existing issues and this hasn't been reported yet
• [x] This is a single bug report (please file separate reports for different bugs)
• [x] I am using the latest version of Claude Code

What's Wrong?

Environment

• Claude Code 2.1.216
• macOS 26.5.2 (build 25F84)
• Plugin marketplace: local-path marketplace (claude plugin marketplace add <local dir>)
• Affected plugins: awesome@2.11.0 (server atlassian-mcp, uvx mcp-atlassian@0.21.1), qa@0.5.0 (server zephyr-mcp, bash-launched bundled Node script)

Summary

Two of seven plugin MCP servers pass claude mcp list's health check ("✔ Connected") but their tools never appear in the live session — not discoverable via tool search, not callable. All other servers in the same plugins register correctly. Reproduced across 6+ separate session restarts / /reload-plugins runs over multiple days (2026-07-03 through 2026-07-21), including reloads performed after the server was already confirmed healthy — so it isn't just "server wasn't up yet at reload time."

Affected servers

• plugin:awesome:atlassian-mcp — mcp__plugin_awesome_atlassian-mcp__* never registers
• plugin:qa:zephyr-mcp — mcp__plugin_qa_zephyr-mcp__* never registers

Working servers for comparison (same plugins, same restarts)

• plugin:awesome:microsoft-mcp (also uvx-launched — rules out "uvx is the problem")
• plugin:awesome:gitlab-mcp, plugin:awesome:tempo-mcp, plugin:awesome:estimate-mcp
• plugin:qa:playwright-test (also bash/node-launched bundled-style — rules out "bundled node server is the problem")

Evidence the broken servers are themselves healthy

Manually ran the exact atlassian-mcp launch command and sent a raw MCP initialize + tools/list request over stdin (bypassing Claude Code entirely). It returned a complete, correctly-schema'd 22-tool list matching the plugin's ENABLED_TOOLS config — even with a deliberately invalid PAT (this server's auth is lazy, only checked on tool invocation). So the server process, its config, and its MCP implementation are all provably correct; something specific to Claude Code's client-side registration for these two servers is dropping their tools.

Ruled out as causes: missing deps (uv/uvx installed correctly), invalid tokens, slow startup (~1.5s, comparable to working microsoft-mcp), oversized tools/list payload (~42KB/22 tools, unremarkable vs. working servers).

Impact

Any workflow depending on these two servers' tools (e.g. reading/writing Jira issues, Zephyr Scale test plans/cycles/cases) is completely blocked, with no visible error — claude mcp list reports success, giving no indication anything is wrong.

What Should Happen?

When claude mcp list reports a plugin's MCP server as "Connected", that server's tools should be discoverable and callable in the same session — the same way the other 5 servers from the same two plugins (microsoft-mcp, gitlab-mcp, tempo-mcp, estimate-mcp, playwright-test) already behave after the same restart/reload.

claude-code-mcp-bug-report-logs.md

Error Messages/Logs

Steps to Reproduce

1. Install a plugin bundling multiple MCP servers, including at least one uvx-launched Python server with required sensitive userConfig tokens.
2. Configure required tokens, /reload-plugins (or restart).
3. claude mcp list → server shows Connected.
4. Search for that server's tools (or try calling one) → not found / not registered, while sibling servers in the same plugin register fine.

Claude Model

Sonnet (default)

Is this a regression?

I don't know

Last Working Version

_No response_

Claude Code Version

2.1.216

Platform

Anthropic API

Operating System

macOS

Terminal/Shell

Terminal.app (macOS)

Additional Information

_No response_