Phoenix
by Arize-ai
Phoenix's MCP server gives an agent read/write access to your LLM traces, prompts, datasets, and experiment results for observability and evaluation work.
Run via npx
npx -y @arizeai/phoenix-mcp@latest --baseUrl https://your-phoenix-host --apiKey your-api-key
About
Arize Phoenix is an open-source LLM observability and evaluation platform, and its @arizeai/phoenix-mcp package puts that platform's data behind an MCP server: an agent connected to it can list and inspect projects, pull traces and spans with their annotations, explore multi-turn conversation sessions, and manage prompts — listing them, fetching versions, tagging them — directly from a chat interface.
It also reaches into the evaluation side of Phoenix: listing datasets and their examples, adding new examples to a dataset, and retrieving experiment results scoped to a given dataset, plus listing the annotation configs a project uses to score its traces. Running it is a single npx command with a Phoenix instance's base URL and API key, either inline as CLI flags or via environment variables, and it drops straight into Claude Desktop or Cursor's MCP config.
Key features
- Lists and inspects Phoenix projects, traces, spans, and span annotations
- Explores multi-turn conversation sessions recorded in Phoenix
- Manages prompts: list, get, version, and tag directly through MCP
- Lists datasets, retrieves their examples, and adds new examples
- Retrieves experiment results scoped to a specific dataset
- Project-scoped operation with an optional default-project setting
Use cases
- Ask an agent to summarize why a specific trace in Phoenix produced a bad response
- Have an agent pull the latest tagged version of a prompt before reusing it elsewhere
- Add a new evaluation example to a dataset directly from a chat session after spotting an edge case
- Compare experiment results across a dataset without leaving your MCP client
Available tools
list traces
Lists recorded LLM traces for a project, optionally with span-level annotations.
get prompt / list prompt versions
Retrieves a managed prompt and its version history, including tags.
list dataset examples / add dataset example
Reads existing examples in a dataset or appends a new one for evaluation.
get experiment results
Retrieves experiment results scoped to a specific dataset.
Frequently asked questions
Do I need a separately hosted Phoenix instance for this to work?
Yes — the MCP server is a client that talks to a running Phoenix instance via its baseUrl and apiKey; it doesn't bundle the observability backend itself.
Can it only read data, or can it also add to my evaluation datasets?
Both — alongside read-oriented tools for traces and prompts, it can add new examples to a dataset directly, which is a write operation.