Catch Breaking SDK Releases & MCP Schema Drift Before Your Agents Fail

Responses API and realtime event payloads updated; Chat Completions tool_choice strict schema enforcement tightened in v7.x.

Major v1.x transition drops legacy completions helpers and enforces typed beta headers for computer-use, code-execution, and MCP connector blocks.

Streamable HTTP transport replaces legacy SSEClientTransport defaults; McpServer.registerTool inputSchema requires Zod object shape or JSON Schema draft-2020-12.

AI SDK v7 unifies LanguageModelV3 provider specifications, replaces maxSteps with stopWhen(stepCountIs(n)), and restructures UIMessage stream parts.

v4 provider targets LanguageModelV3 spec and defaults to OpenAI Responses API for reasoning and built-in web/file search tools.

v4 provider requires AI SDK 7+ and moves thinking budget configuration under providerOptions.anthropic.thinking.

v1.x removes deprecated RunnableBinding legacy call signatures and standardizes ToolMessage artifact and content_blocks metadata.

v1.x StateGraph Annotation reducer syntax and interrupt() resume semantics updated; createReactAgent moved to @langchain/langgraph/prebuilt.

v2.0 requires @langchain/core >= 1.0 and @modelcontextprotocol/sdk >= 1.20; MultiServerMCPClient transport config enforces explicit transport type.

Rapid v1.x cadence updates Agent memory thread scoping, createTool inputSchema validation, and workflow step execution context.

0.x minor releases regularly refine handoff filter signatures, hosted MCP tool definitions, and RealtimeAgent voice session events.

Replaces deprecated @google/generative-ai package; v2.x standardizes ai.models.generateContent and functionDeclarations schema typing.

v4.x overhauls stagehand.act() and stagehand.extract() options object, separating modelClient configuration from Playwright page lifecycle.

v3.x generated via Speakeasy updates camelCase request fields and agents.complete tool call response union types.

v8.x defaults to Cohere Chat v2 endpoint (CohereClientV2) with OpenAI-style messages and strict tool parameter schemas.

v3.x Python SDK enforces Pydantic v2 strict models on client.responses.parse and removes legacy beta.threads assistants helpers.

v2.x drops Python 3.9 support, requires typed thinking blocks on multi-turn tool loops, and updates streaming citation events.

v2.x restructures FastMCP lifespan context, StreamableHTTP session handling, and CallToolResult structuredContent validation.

FastMCP v4.x introduces OpenAPI/MCP composition mounts, OAuth proxy providers, and strict output_schema serialization.

v1.x standardizes human-in-the-loop interrupt() and Command(resume=...) primitives and moves prebuilt agents to langgraph.prebuilt.

v1.x Flows and Crew memory backends use native LiteLLM / provider clients and enforce BaseTool args_schema Pydantic v2 classes.

v2.x replaces result_type with output_type on Agent(), and renames result.data to result.output across all run methods.

High-frequency releases update model cost maps, Anthropic/Bedrock tool call argument normalization, and MCP gateway bridge routes.

DSPy v3.x unifies dspy.LM() with LiteLLM identifiers, removes legacy dspy.OpenAI classes, and updates dspy.ReAct tool wrapping.

Updates MCPServerStreamableHttp params, RunConfig tracing processors, and strict JSON Schema validation for @function_tool.

Major v2.0.0 release enforces roots protocol capabilities, tightens symlink path traversal validation, and updates edit_file dryRun diff schema.

Major v2.0.0 release updates MEMORY_FILE_PATH env handling and enforces strict entityType and observations array validation in create_entities.

Major v2.0.0 release adds structured outputSchema validation to sequentialthinking tool responses and requires strict boolean nextThoughtNeeded.

Major v2.0.0 release aligns test tools, prompts, and resources with MCP 2025-11-25 / 2026 spec features including elicitation and audio blocks.

Archived on npm at v2025.4.8 in favor of GitHub's official remote/Docker ghcr.io/github/github-mcp-server; tool names and pagination parameters differ.

Frozen at v0.6.2 in archived reference servers; lacks MCP 2025+ Streamable HTTP and parameterized query tools.

v2025.4.25 requires SLACK_TEAM_ID and SLACK_BOT_TOKEN env vars; legacy workspace token auth fails.

v0.6.2 reference server superseded by Brave's official @brave/brave-search-mcp-server with news/video/image tools.

Archived at v2025.5.12; teams migrating to @playwright/mcp or chrome-devtools-mcp for accessibility snapshots.

v2025.1.14 requires local OAuth credentials file path via GDRIVE_CREDENTIALS_PATH.

v0.6.2 requires GOOGLE_MAPS_API_KEY with Places API (New) and Directions API enabled.

v2025.4.25 transitions to official @redis/mcp-server for Redis Streams, hashes, and vector index tools.

Superseded by awslabs.bedrock-kb-retrieval-mcp-server with reranking and metadata filters.

v2.x requires session token authentication by default on proxy server port 6277 to prevent DNS rebinding / local RCE.

v0.3.x scopes tools by resource permission flags (--tools=all or granular resource.action) and recommends remote https://mcp.stripe.com.

Rapid 0.0.x releases rename vision mode flags (--caps=vision,pdf) and update browser_snapshot element ref identifiers.

v0.14.x supports --read-only and --project-ref scoping; execute_sql and apply_migration split DDL vs DML operations.

Local CLI server supplemented by dedicated remote OAuth MCP servers (workers-bindings, observability, radar).

v2.x replaces raw Notion-Version header config with simplified markdown block append payloads and OPENAPI_MCP_HEADERS env format.

v0.42.x consolidates search_issues and get_issue_details parameters, requiring organizationSlug on all multi-org tokens.

v0.19.x adds dynamic actor discovery tools (search-actors, add-actor) instead of preloading all actor schemas at startup.

v3.x upgrades to Firecrawl v2 API: firecrawl_extract requires schema or prompt, and firecrawl_scrape removes deprecated pageOptions.

v3.x enables remote hosted MCP by default and renames tool identifiers to web_search_exa and deep_researcher_start.