New
- CLI JS hooks — Claude Code, Codex CLI, and Gemini CLI integrations now use lightweight Node.js hooks instead of Python. 25KB per hook, no Python dependency needed. Send traces via OTLP v2 JSON.
- OpenAI Agents SDK (TypeScript) — auto-instrument agent, tool, generation, and handoff spans.
- Respan CLI —
@respan/cli v0.6.0 with JS hooks, --api-key auth default, and respan integrate for all 4 coding agents.
- 95+ integration overview — searchable integration page with filter pills for Agent Frameworks, LLM SDKs, Coding Agents, Vector Databases, and Model Providers.
- Monitor templates — automation template gallery with drag-and-drop workflow builder.
Improved
- OpenTelemetry integration page rewritten with full semantic convention mapping (GenAI, Traceloop, OpenInference)
- JS tracing SDK reliability improvements
- Docs migrated to new documentation platform
New
- GPT-5.4 model support — added gpt-5.4, gpt-5.4-mini, and gpt-5.4-nano models.
- Experiment seed field — set a seed for reproducible experiment runs.
- Dataset outcome spans — dataset evaluation tasks now emit outcome spans for observability.
- Payment methods management — manage payment methods directly from the billing page.
- Onboarding flow — improved onboarding with integration setup and Playwright-tested auth flows.
Improved
- Service tier promoted from metadata to top-level column
- Auto-create webhook mappings for request log events
- Filter infrastructure refactored for type safety and OpenAPI alignment
- Agent tool error responses sanitized
New
- OpenAI SDK TypeScript integration — auto-instrument OpenAI calls with
@respan/instrumentation-openai. Includes streaming, tool calls, multi-turn, structured output, and batch API examples.
- is_cached_by_model — new cache option to invalidate caches when switching models. Set
cache_options.is_cached_by_model: true.
- Span pinning — pin spans for infinite retention with
is_pinned: true. PATCH existing spans to pin them.
Improved
- Customer identifier now has 254 character limit with auto-truncation
- Properties field added for typed metadata (preserves native types unlike metadata)
New
- Claude Agent SDK integration — full tracing support for Anthropic’s Claude Agent SDK with auto-instrumented spans.
- Unified Python & TypeScript SDKs —
respan package with plugin system, propagate_attributes(), and log_batch_results().
- Pydantic AI integration — trace Pydantic AI agent runs with auto-instrumentation.
- Vercel AI SDK updates — provider tabs and gpt-5-mini support.
Improved
- RAG pipeline tracing cookbook added
- Updated docs URL structure with flat integration paths
New
- Prompt schema V2 — new merge strategy where prompt config always wins. Use
schema_version: 2 for all new integrations. Includes patch parameter for runtime overrides.
- Fallback models — add
fallback_models to any gateway request. The gateway automatically tries each model in order if the primary fails.
- Anthropic Agents SDK integration — trace agent runs, tool calls, and LLM generations for both Python and TypeScript.
- RubyLLM gateway integration — route RubyLLM requests through the Respan gateway.
- Enterprise MCP — dedicated endpoint at
mcp.respan.ai/mcp/enterprise with OAuth, JWT, and API key auth.
Improved
- Prompt version API endpoints cleaned up
- Trace delete endpoint added
New
- Agent chat — conversational interface for creating and managing automations. Multi-agent architecture with specialized sub-agents for log analysis, evaluation, and prompt engineering.
- Evaluation workflows — run evaluators as part of automation workflows with data source resolution and sampling.
- Responses API support —
customer_credentials passthrough now works with OpenAI’s Responses API.
- Automation monitoring — conditions with aggregation metrics, throttling, and data source configuration.
Improved
- Dashboard analytics with 12 new MCP tools
- Evaluator filter definitions with dynamic choices
New
- Rebrand: Keywords AI to Respan - all URLs, packages, and APIs updated. Legacy endpoints redirect automatically.
- Platform MCP server - access logs, traces, prompts, and customer data from AI coding tools via OAuth, JWT, or API key. Supports Claude Code, Cursor, and Codex CLI.
- Docs MCP server - search Respan documentation from your editor. No API key needed.
- Bedrock bearer token auth - AWS Bedrock now supports API key authentication in addition to IAM.
- Span comparison keys - compare spans side-by-side with dedicated comparison keys.
- PostHog integration - send LLM usage data to PostHog for product analytics.
Improved
- Docker Compose dev environment with hot-reload for runserver, worker, and CH writer
- Webhook delivery tracing for dogfood observability
New
- Reasoning tokens tracking — track reasoning tokens separately for o3, gpt-5, and other reasoning models. Shows as a dedicated field alongside prompt and completion tokens.
- Export pipeline — partition-based export with sampling percentage and field selection.
- MCP tools expansion — added tools for credit transactions, API key management, experiments, scores, traces, automations, and dashboard analytics.
- Rate limiting — global rate limiting across all API endpoints for stability.
Improved
- Prompt cache hit tokens bug fix
- Dead letter queue (DLQ) observability and retry tools
New
- Models summary endpoint — aggregate model usage statistics via API.
- MCP tools for prompts — create and manage prompt versions directly from Claude Code or Cursor.
- Experiment V2 eval results — run evaluations and view results directly in experiment tables.
Improved
- Automation workflows now support name, description, and throttling
- Streaming logging improvements
New
- Views page — save and star filtered views across Logs, Traces, Users, and Prompts. Add descriptions and filter by creator.
- Auto top-off — automatically recharge credits when balance drops low.
- Webhook secret signing — webhooks now use
whsec_ signed secrets for verification.
- Batch API — submit large batches of requests for async processing at 50% cost.
Improved
- Better custom provider management
- Saved filters now support POST-body filtering
