NewYour coding agent can read the release notes before it upgrades.Set up the MCP server →
PyPI
PyPI, #4973 most downloaded on PyPI
AI Observability and Evaluation
Latest stable
21.0.0
Last release 23 days ago
30 Oct 2026
Ships on a steady schedule
a new release about every 8 days
Nearly every release is documented
notes for 60 of the last 60 stable releases
19 versions withdrawn
withdrawn after publishing
4 years old
714 releases · first in 2023
Drag across the ruler to compare, or tap a release to open it.
annotations: add DECISION and AGENT kinds
Your coding agent can read these notes before it upgrades. Set up the MCP server →
api: expose prompt version creation timestamps
playground: refresh MiniMax text model presets
factory: link session IDs in factory comments to the Claude Console
factory: scaffold managed agents issue-to-PR factory
deps: require Node releases patched in the October 2026 security release
server: export server logs over OTLP
playground: add Meta muse-wonder-1.0 and gemini-4-argon
ui: keep labels on Escape and fix table sort, a11y and accessor warnings
20.21.3 (2026-10-13) Documentation fix dead Read the Docs links
experiments: coerce average repeated-run latency to float on PostgreSQL
agents: upgrade pydantic-ai for CVE-2026-107286
result (#16902) (d927595)harbor: prompt hill-climb task on a text-to-SQL dataset
datagen: send each recorded application to its own project
tracing: show eval results in the trace tree
agents: mark bash tool spans as errors on non-zero exit codes
add gpt-6-sol, gpt-6-luna, and claude-opus-5-5 to the cost tables and the playground
agents: mount configured skills in MCP and PXI
datasets: editable dataset examples table
annotations: add GraphQL mutations for span, trace, and session notes
add Claude Code plugin and marketplace
client: add ATIF tracing to Harbor plugin
add MiniMax provider integration
server: move the MCP server to FastMCP 4 and the stateless protocol
server: replace deprecated json_encoders with an IsoDatetime serializer
add skill tools to mcp and use them in pxi
server: GET /v1/model_providers no longer returns custom providers or a next_cursor, and built-in entries expose provider rather than kind + provider_
next_cursor, and built-in entries expose provider rather than kind + provider_key. Custom providers move to GET /v1/custom_model_providers. The endpoint is unreleased, so no published client is affected.## 20.3.0 (2026-08-16) ### Features * traces: expression filter DSL
cost: update built-in model token prices
agents: emit a machine-readable approval decision on gated PXI tools
auth: support JWT client assertion for OAuth2 login
add semantic coloring to project metric tooltips
api: add REST endpoints for setting tags on experiments
add copy actions to session turns
app: make the Mastra snippet's OTLP path append idempotent
ui: collapse all but the last message in LLM spans
cli: make px setup's trace verdict the run's result (#14904) (253cfcf), closes #14876
server: OpenAI-compatible v1/chat/completions proxy with server-side credentials
add project evaluation metrics charts
add experiment evaluation metrics charts
agent: add Claude Opus 5 to the curated model list
client: add span_ids filter to get_spans in python and typescript clients
add evaluation metrics aggregation APIs
components: align span kind icon and timeline bar colors with token treatment
auth: negotiate HTTP/2 for OAuth2/OIDC provider requests
cli: improve server startup banner
cli: px setup mcp — register the Phoenix remote MCP server with a coding agent
api: REST CRUD for user and system API keys
sessions: use interval-overlap semantics for session time-range filters
agents: force tracing with debug env
One column per quarter.