Skip to main content
This release has two practical goals: make server-side code evaluators easier to run for teams that already use Docker, and help MCP-connected agents answer deeper questions about Phoenix data. You can authenticate once in Phoenix, run Python and TypeScript evaluators in Docker’s hosted sandbox environment, and keep Phoenix controls for cleanup and outbound network access. MCP clients also get GraphQL schema discovery, read-only queries, and permitted mutations, alongside Claude Haiku 5.5 and Mistral cost tracking.

​
Docker Sandboxes for Code Evaluators

October 8, 2026 Available in arize-phoenix 20.20.0+ Docker Sandboxes gives code evaluators a hosted runtime when your team already has Docker Agentic Platform access. Add your Docker credentials in Settings → Sandboxes, then choose Docker Sandboxes on the sandbox configuration you use for Python or TypeScript evaluators. Phoenix still owns the evaluator flow: it sends the run to Docker, applies the network setting, and cleans up when the run finishes.
Phoenix Settings Sandboxes page with the Docker Sandboxes credentials modal showing Docker ID and Docker personal access token fields

Configure Docker Sandboxes once, then use it for Python and TypeScript code evaluators

  • Run both evaluator languages. Python runs on python:3.13-slim and TypeScript on node:24-slim.
  • Authenticate once. Add your Docker ID as DOCKER_ID and a personal access token with the sandbox:use permission as DOCKER_PAT, either in Settings or as server environment variables.
  • Control network access. Deny attaches a phoenix-deny-all-egress network policy that stops sandbox traffic from leaving the environment.
  • Clean up automatically. Phoenix deletes each sandbox when the evaluator finishes. Docker also expires every sandbox after 10 minutes.

Docker Sandboxes

Requirements, setup, and sandbox behavior

Code Evaluators

Write Python and TypeScript evaluators that run on the server

​
Claude Haiku 5.5 and Mistral Cost Tracking

October 8 to 11, 2026 Available in arize-phoenix 20.20.0+ (Claude Haiku 5.5) and 20.21.0+ (Mistral pricing) You can test Claude Haiku 5.5 in the playground and see token cost on Mistral spans without adding custom pricing.
  • Claude Haiku 5.5 is in the playground for the Anthropic and Amazon Bedrock providers, with adaptive thinking support.
  • Retired Anthropic models no longer appear in model lists. Phoenix removes Claude Opus 4.1, Sonnet 3.7, Sonnet 3.5, Haiku 3.5, and Haiku 3.
  • Mistral AI calls now show a cost. Phoenix prices Mistral spans from the built-in model price list, the same way it handles other providers.

Cost Tracking

How Phoenix prices tokens on your spans

​
MCP Server: GraphQL Tools and More SQL Tables

October 8, 2026 Available in arize-phoenix 20.20.0+ Use GraphQL when an MCP client needs structured Phoenix data that REST does not expose. The agent can inspect the schema and run authorized queries from the MCP server, instead of falling back to shell access or one-off scripts.
  • Discover and query the schema. describeGraphqlSchema searches by free text or exact type names, and executeGraphqlQuery runs read-only queries with the caller’s permissions.
  • Run permitted mutations when writes are enabled. executeGraphqlMutation can run any mutation the caller’s role allows. Phoenix does not expose this tool on read-only deployments.
  • Ask SQL questions over more Phoenix tables. executeSql can query trace annotations, evaluators, prompts and prompt versions, dataset splits and labels, experiment tags, jobs and logs, sandbox configs, and trace retention policies.
  • PXI uses analytics SQL for data questions, so it spends fewer turns answering aggregate questions about your traces.

MCP Server

Connect an MCP client and see the full tool list

​
Additional Improvements

October 8 to 13, 2026 Available in arize-phoenix 20.20.0–20.21.3, arize-phoenix-evals 3.9.1+, and @arizeai/phoenix-client 7.16.1+
  • Evaluator metadata stays attached. The server keeps metadata nested under result, and the TypeScript client sends metadata at the top level.
  • Stale input mappings no longer break evaluator runs when you remove a variable from the template or evaluate() signature.
  • Experiment resume cleans up replaced runs by dropping evaluations from an errored run when a new run takes its place.
  • Experiment metric comparisons are consistent because Phoenix measures the base experiment and compare experiment the same way.
  • Span ingestion recovers from database failures more reliably and reports failed spans and the last successful insert as Prometheus metrics.
  • Projects from span ingestion are easier to scan because Phoenix gives each new project a random gradient.
  • arize-phoenix installs fewer heavy dependencies because SciPy and scikit-learn are no longer included by default.
  • Synchronous eval runs retry rate-limit errors instead of failing the whole run on the first 429.