v1.6.1

Agents SDK — Release Notes v1.6.1

New Tools

4 new generalist tools added to the standard tool registry:

• Calculator — evaluates mathematical expressions via embedded Python
• Weather — current conditions lookup by location
• Datetime — current date/time with timezone support
• Fetch Webpage — retrieves and strips raw page content for agent consumption

Brave Search replaces SerpAPI as the default web search backend — significantly faster (~100–300ms vs ~500–1500ms), direct index with no proxy layer, and no redundant LLM summarization pass.

OpenClaw Feature Parity

Full OpenClaw-style workspace support is now built into the core SDK:

• loadWorkspaceConfig(dir) — single call loads persona files (AGENTS.md, SOUL.md, USER.md, MEMORY.md, TOOLS.md, IDENTITY.md), MCP servers, and skills automatically
• SkillLoader — loadFromDefaultPaths() scans ~/.agents/skills, ~/.claude/skills, skills/, .agents/skills/ automatically; supports requires.env frontmatter for dependency
  validation
• Memory improvements — BM25 scoring fixes (word-boundary regex, zero-score filtering, decay as multiplier), flushToPersistent with optional LLM for agentic fact
  extraction before compaction
• OpenClaw demo — full reference implementation at demos/openclaw/

In-Car AI Demo Refresh

• Restructured to OpenClaw pattern with per-demo .agents/ directory (persona files + MCP config)
• HTTP server extracted to ui.hpp — single-header, no separate compilation unit
• Apple Music MCP moved from hardcoded C++ to demos/in-car/.agents/mcp.json

Windows Compatibility

• Python tool: replaced popen/pclose/mkstemp with Windows equivalents (_popen, _pclose, GetTempPathA/GetTempFileNameA)
• Calculator tool: switched from shell -c quoting (broken on Windows cmd) to temp file approach
• SkillLoader: checkRequirements uses where on Windows instead of which

v1.5.1

• Bug fix for streaming in all LLM providers.
• Migrated OpenAI LLM provider to use the responses API.
• Improvements and fixes to Websearch tool.
• Internal improvements to autonomous agent.
• UI & runtime Improvements to demos.

v1.4.7-1

• Hotfix to v1.4.7
• Implemented new local, offline Voice Agent.
• Integrated into Edge AI STT and TTS voice stack.
• Introduced demos showcasing possible integrations.
• Bug fixes and internal API cleanup.

v1.4.7

• Implemented new local, offline Voice Agent.
• Integrated into Edge AI STT and TTS voice stack.
• Introduced demos showcasing possible integrations.
• Bug fixes and internal API cleanup.

v1.4.1

• Internal API cleanup for consistency
• Addressed various bugs in LLM providers.
• Reworked ReAct planning for better performance.
• Added MCP support with auto-discovery [Only available in Pro version].
• Implemented Voice agent & added STT, TTS, & VAD support through Edge AI provider [Only available in Pro version].

v1.3.0

• Validated Windows support
• Notable API cleanup
• Bug fixes to Anthropic LLM provider interface for 4.x Claude model support
• Added new respond tool

v1.1.0

• Migrated from CPR library to cpp-httplib for HTTP usage
• Added support for document understanding (see multimodal_example)
• Various bug fixes in edge conditions
• Consistency updates to examples

v1.0.0

• Initial production release running.
• Documentation generated and posted on website.
• Web search, file read, file write, summarization, wikipedia, and python execution tools developed.
• Multiple examples developed to demo tool calling, different workflows, and agents available.