Tool Use

Google ADK Python v1.23.0 introduces OpenTelemetry-based tracing for BigQuery, automatic session creation, and enhanced tool support with JSON schema declarations. The release includes 30+ features and bug fixes across core, tools, evals, observability, and integrations, with a breaking change replacing custom ContextVar implementation with OpenTelemetry.

Google ADK Python v1.24.0 introduces a breaking change to credential manager API (accepting `tool_context` instead of `callback_context`) and delivers significant enhancements across web UI, core functionality, tooling, and observability. Key features include A2UI v0.8 integration for native UI component rendering, toolset authentication framework, asynchronous OpenAPI tools, improved accessibility with arrow-key navigation, and numerous bug fixes for session management, credential handling, and deployment validation.

Google ADK Python v1.25.0 release introduces significant enhancements including a simple prompt optimizer demo, new CLI flags for auto session creation, event-delta memory management, and post-invocation token compaction. The release adds support for external access tokens in Google credentials, new tools like MCP resource loading and SkillToolset, plus health/version endpoints to the web server. Multiple bug fixes address async iteration, event loop issues, metadata preservation in streaming responses, and PostgreSQL timestamp handling.

Google ADK Python v1.26.0 release introduces major features including intra-invocation token compaction, Apigee LLM chat completions integration, Agent Skills spec compliance, BigQuery search catalog support, and an Agent Registry system. The release also includes numerous bug fixes for streaming tools, memory services, authentication, and various integrations across core, models, tools, evals, and services components.

Release 2025.11.25 of modelcontextprotocol/servers updates five packages including sequential-thinking, everything, filesystem, memory, and git servers. All packages are bumped to version 2025.11.25.

OpenClaw v2026.2.22 release adds Mistral provider support with memory embeddings and voice, introduces auto-updater functionality with dry-run preview, and expands channel integrations with native Synology Chat support. The release includes significant improvements to memory search with multilingual stop-word filtering (Spanish, Portuguese, Japanese, Korean, Arabic), enhanced Slack threading and media handling, and numerous bug fixes across Discord voice, Docker setup, webchat performance, and browser extension relay stability.

OpenClaw v2026.2.23 release adds first-class Kilo Gateway and Moonshot provider support, introduces prompt caching documentation and per-agent parameter overrides, and hardens session maintenance with disk-budget controls and security headers. The release includes numerous fixes for browser SSRF policies, Telegram polling/reactions, context overflow detection, and agent reasoning/compaction behavior.

OpenClaw v2026.2.25-beta.1 release includes improvements to Android chat streaming and startup performance, UI enhancements for mobile compose actions, and a major refactor of heartbeat configuration policy. The release addresses numerous critical fixes across Telegram, Slack, Discord, and LINE channels, including subagent delivery state machine improvements, webhook handling, session threading, media routing, and typing indicator management.

OpenClaw v2026.2.25 release includes improvements to Android chat streaming and startup performance, UI enhancements for mobile compose actions, and a breaking change to heartbeat direct message delivery policy (now `allow` by default). The release contains numerous bug fixes across Telegram, Slack, Discord, LINE channels, agent routing, cron scheduling, and model fallback logic.

OpenClaw v2026.2.26 introduces major features including External Secrets Management with a complete workflow, ACP/Thread-bound agents as first-class runtimes, new agent routing CLI commands for account-scoped management, and Android device capabilities. The release also includes numerous fixes for Telegram DM allowlist inheritance, delivery queue recovery, typing indicators across channels, and various platform-specific lifecycle and authentication improvements.

OpenClaw v2026.3.1 release introduces major enhancements across multiple platforms including adaptive thinking defaults for Claude 4.6, Kubernetes health check endpoints, expanded Android device capabilities (camera, notifications, sensors), Discord/Telegram session management improvements, and WebSocket-first OpenAI streaming. The release also adds localization support (German, zh-CN), Feishu document/chat tooling, LanceDB memory customization, and numerous reliability fixes for Android, Windows, LINE, and Slack integrations.

OpenClaw v2026.3.2-beta.1 introduces major features including expanded SecretRef support across 64 targets, native PDF analysis tools, enhanced outbound adapters with multi-media support, and improved Telegram streaming with DM-specific optimizations. The release includes breaking changes to onboarding defaults, ACP dispatch behavior, and plugin SDK HTTP handler registration, alongside numerous fixes for plugin command validation, Telegram token handling, and group system prompts.

OpenClaw v2026.3.2 introduces major features including expanded SecretRef support across 64 targets, native PDF analysis tools, enhanced outbound adapters with multi-media support, and improved session attachments. The release includes several breaking changes around default tool profiles, ACP dispatch enablement, and plugin SDK HTTP handler registration, along with numerous fixes for Telegram, Discord, Feishu, and other channel integrations.

gpt-code-search is an open-source tool that enables natural language code search across any codebase using OpenAI's GPT-4 and function calling. It runs locally without requiring code indexing or repository uploads, offering functions like search_codebase, get_file_tree, and get_file_contents to help the LLM retrieve and analyze code context.

Strata is an open-source MCP server by Klavis AI that enables AI agents to access thousands of API tools through progressive discovery rather than overwhelming them with all options at once. The system guides agents through hierarchical categories and actions, achieving 15.2% higher accuracy than official GitHub and Notion servers while supporting complex multi-app workflows with 83%+ accuracy.

A custom MCP (Model Context Protocol) server that connects Blender to LLMs like ChatGPT and Claude, enabling natural language control of 3D scene generation. Users can describe complex environments (villages, landscapes, animations) in plain text and the system automatically builds them in Blender with support for spatial reasoning, iterative editing, and object hierarchy management.

Wild Moose is an autonomous AI agent designed for production debugging that connects to observability data like logs and metrics from Datadog. Rather than focusing on code generation, it performs root-cause analysis by exploring multiple data sources through conversational interfaces, combining code execution with API invocations to answer debugging questions.

Nous is an open-source TypeScript agent framework combining features of CrewAI, OpenDevon, and LangFuse, designed for building autonomous and software engineering agents with integrated tooling. It includes a Web UI, database persistence, tracing, human-in-the-loop functionality, and a novel autonomous agent that generates Python code executed in a WebAssembly sandbox for improved reasoning and reduced LLM costs.

Pomerium introduces an Agentic Access Gateway, an open-source feature that extends their zero-trust identity proxy to secure AI agents with fine-grained, context-aware authorization. The solution provides centralized policy enforcement, just-in-time credentials, and audit trails for AI agents accessing internal apps and APIs, treating agents as first-class identities rather than relying on static OAuth scopes.

BrowserOS is an open-source Chromium fork enabling non-developers to create and run browser agents locally. After testing three UX approaches (drag-and-drop workflows, one-shot agents, and plan-follower agents), the team found that having users provide simple natural language plans while the LLM executes each step achieved the best balance—improving success rates from 30% to ~80% even with local models like those run via Ollama or LMStudio.