Arize Phoenix adds Claude Code, Codex and Cursor plugins

Arize released Phoenix 20.10.0 on September 11. It adds a Claude Code plugin and marketplace, an installable Codex plugin, and a Cursor plugin for Phoenix's remote MCP server and public skills.

The bigger story sits in the next day's release. Phoenix 20.11.0 fixed a bug where the API key passed to mcp-remote sat in the process argv, visible in the process list to anyone else with access to the machine, per Phoenix's 20.11.0 release notes. That's a credential leak on any shared box or CI runner, not a theoretical one.

The same release also serves a shared skills root on the /mcp mount and adds phoenix-error-analysis, a skill for triaging failures inside the tool. Three coding assistants, one observability platform, wired through the same protocol across two consecutive releases. The earlier release added span-level cost filtering and a completeness evaluator too, so the plugin push wasn't the only thing moving.

Argv secrets are an old bug class. Command-line arguments are visible to anyone who can list processes on the box, which is why secrets belong in environment variables or config files instead. Phoenix isn't the first tool to make this mistake and won't be the last.

As MCP servers multiply and every editor wants a plugin for every observability tool, the connection code linking them keeps getting written from scratch. Anyone standing up an MCP integration that hands a client an API key should check where that key ends up when the process spawns. argv is the wrong answer.