Skip to content
Agents tracked: 283 Downloads (7d): 239M down 5.7% GitHub stars: 6.2M VS Code installs: 151M Releases (7d): 338 Agent status: 1 with issues Updated Oct 9, 2026
AgentGid

Phoenix changelog: what's new each month

Every stable Phoenix release summarised by month: the highlights, new features, improvements, fixes and anything you need to act on. 3 months covered; the current month updates daily.

October 2026 · so far

3 releases: 20.19.0 → 20.20.0

October brought improvements to data generation and tracing infrastructure, along with new capabilities for SQL handling and sandbox environments. The month focused on better resource management and expanded tooling options.

Highlights

  • New Docker Sandboxes provider for running isolated environments.
  • Added GraphQL schema, query, and mutation tools to MCP for broader data access.
  • Introduced DECISION span kind across the tracing platform for better visibility.
  • Expanded SQL table allowlisting in MCP to include annotation, evaluator, prompt, split, and job tables.
  • Added claude-haiku-5-5 model support and removed retired Anthropic models from the playground.

New

  • Docker Sandboxes provider for isolated execution environments.
  • GraphQL schema, query, and mutation tools in MCP.
  • DECISION span kind added to tracing platform.
  • Prompt hill-climb task for Harbor on text-to-SQL datasets.
  • Configurable MCP code mode for Harbor benchmarks.

Improved

  • Data generation now sends each recorded application to its own project.
  • External resources flag now honored for WASM and UI components.
  • Evaluation runs now retry on RateLimitError instead of failing completely.
  • MCP/SQL now allows access to annotation, evaluator, prompt, split, and job tables.
  • Upgraded pydantic-ai to version 2.52 with improved token handling.

Fixed

  • Dropped Anthropic maxtokens override in agents for better compatibility.
  • Separated harbor-verifiers from root project dependencies to reduce bloat.

September 2026

58 releases: 7.7.0 → 4.3.15

September brought expansions to Phoenix's model provider ecosystem, client libraries, and evaluation capabilities. The month saw numerous improvements to tracing, data management, and agent tooling.

Action needed
  • GET /v1/modelproviders no longer returns custom providers or nextCursor; built-in entries now expose provider instead of kind and providerKey.
  • rootSpansOnly GraphQL parameter removed in favor of span filter DSL.
  • Project.traceAnnotationsNames renamed to traceAnnotationNames.
  • Individual trace error and latency parameters deprecated in favor of filter expressions (existing behavior retained).

Highlights

  • Added support for multiple new model providers including Z.ai, Meta Muse Spark, and MiniMax, plus new models like Claude Opus 5.5 and GPT-6 variants.
  • Expanded dataset management with editable examples tables, dataset splits API, and new helpers for creating and managing splits.
  • Enhanced trace filtering and querying with new error, latency, and filter expression support across trace and session endpoints.
  • Added comprehensive evaluation tools including completeness, retrieval relevance, and PII detection evaluators.
  • Introduced secrets management and prompt version tagging helpers for better configuration management.

New

  • Project retention policy assignment and trace transfer capabilities between projects.
  • getCurrentUser typed helper and secrets management via upsertOrDeleteSecrets.
  • Dataset split creation, updating, and deletion helpers with filtering support.
  • Prompt version metadata support and prompt-scoped version tagging helpers.
  • Span, trace, and session annotation mutations via GraphQL.
  • Completeness evaluator to judge whether active user requests in conversations were completed.
  • Retrieval relevance evaluator for evaluating search and retrieval quality.
  • Bash tool integration for agents with error detection on non-zero exit codes.

Improved

  • OpenAI SDK peer dependency range widened to support both v6.10.0+ and v7.0.0 without conflicts.
  • Session objects now expose cumulative prompt, completion, and total token counts.
  • Token counting improved to count only leaf LLM spans in project totals.
  • Trace list endpoint now sorts by start time for better organization.
  • Filter expressions added to getTraces and listSessions for more flexible querying.
  • Classification evaluators now accept AI SDK evaluation models like TypeSafe's Jev.
  • PXI slash-command hints now support arrow key navigation, Tab completion, and highlighted selection.
  • Model token pricing tables continuously updated throughout the month for latest models.

Fixed

  • Memory-safety and correctness bugs fixed in phoenix-sqlean C driver.
  • OpenTelemetry register() crash with opentelemetry-exporter-otlp-proto-http 1.45 resolved.
  • GitHub API calls no longer retry after authorization failures.
  • Experiment JSON/CSV export no longer returns 500 when a run errored.
  • SQLAlchemy 2.1 compatibility added for SQLite support.
  • Annotation configuration validation now checks all configured ID columns, not just the first.

August 2026

19 releases: 19.20.0 → 4.3.4

August brought major updates to Phoenix with new agent capabilities, improved filtering and tracing infrastructure, and expanded REST API endpoints. The month included several releases across core components and client libraries.

Action needed
  • GET /v1/modelproviders endpoint behavior changed: no longer returns custom providers or nextcursor, and built-in entries now expose provider instead of kind and providerkey.
  • Custom providers moved to a different endpoint.

Highlights

  • Agent session persistence now available for storing and managing agent conversation state.
  • New expression filter DSL for traces enables more powerful and flexible query capabilities.
  • Extended REST API with endpoints for dataset splits, experiment tags, API key management, and prompt metadata.
  • Added retrieval relevance evaluator to the evals toolkit for measuring search quality.
  • Improved agent tooling with in-process Phoenix MCP toolset and browser-action meta-tools.

New

  • Copy actions for session turns to easily share conversation content.
  • AI query capability for session filters to find sessions using natural language.
  • Expression filter DSL for traces with comprehensive reference documentation.
  • Retrieval relevance evaluator in evals for assessing search result quality.
  • In-process Phoenix MCP toolset for agents.
  • Browser-action meta-tools replacing client-action tools in PXI.
  • Session trace view accessible via double-click on agent session turns.
  • deletePrompt helper in the prompts API for prompt management.

Improved

  • Project navigation no longer blocks during loader fetches.
  • Evaluation metric chart layout and controls refined for better usability.
  • Span filter DSL now supports trace annotations.
  • Google error classification improved and release versioning corrected.
  • Built-in model token prices updated.
  • Session ID collision prevention for exported agent sessions.
  • Phoenix GraphQL mutations enabled by default in manual approval mode.
  • OpenAI reasoning models now routed to Responses API client.

Fixed

  • Malformed Monty error payloads in MCP are now properly contained.
  • Tool attributes that were rebuilt as objects are now correctly stringified.
  • Object-typed parameters fixture casting to satisfy TypeScript.
  • Object jsonschema coercion in getLLMAttributes toolSchemas.
  • Default processor removal only occurs when actually replacing it.

Summaries are written automatically from the official release notes (full changelog ↗); check the original notes before relying on a detail. Phoenix: pricing, features and alternatives · All changelogs