v.1.3.1

Highlights

• New sandbox backends for SandboxShellTool and SandboxCodeTool: E2BEnvironment (ag2.extensions.e2b) runs them in E2B cloud sandboxes, and ag2.extensions.boxd runs them on boxd Linux microVMs.
• MCP Apps work with browser-based hosts: /mcp no longer redirects POST and the CORS preflight, and the docs include a CORS recipe. Misuse is reported with clear errors, and App documents built on the upstream ext-apps SDK are served as written.
• Extended prompt-cache writes are reported separately: Usage.cache_creation_extended_input_tokens counts OpenAI 12h and Anthropic 1h writes, which are priced differently. The field also flows through telemetry, eval traces and MetricsMiddleware.
• XAIConfig accepts safety_identifier and guidelines.
• Fixes for the AG-UI final reply after ToolResult(..., final=True), OpenAI-compatible streamed tool calls without an index, Gemini 3 thought_signature errors, recursive find_files patterns on Python 3.11–3.12, and MCP App documents read in a non-UTF-8 locale.
• ag2.eval, ag2/agent.py, ag2/extensions and most core modules are now checked with strict mypy.

Upgrade notes

• A2UI v1.0 now speaks the stable upstream spec. 1.3.0 targeted a release candidate. v1.0 messages carry "version": "1.0" instead of "v1.0", the basic catalog requires TextField.value, and custom v1.0 catalogs with $ref targets the spec does not allow are rejected at load. protocol_version="v1.0" and the v0.9 default are unchanged. v0.9.1 client capabilities are now read from the "v0.9" key, so conformant clients' capabilities are no longer dropped.
• MCP-UI is documented as legacy. Use MCP Apps for new work. ag2.mcp_ui is not deprecated and keeps working for hosts that render only MCP-UI resources.

What's Changed

• docs: explain Agent Status Bar with practical examples by @Lancetnik in #3583
• feat(xai): support safety_identifier and guidelines by @vvlrff in #3598
• feat(usage): report extended prompt-cache writes (OpenAI 12h, Anthrop… by @vvlrff in #3599
• fix(usage): carry extended cache writes through telemetry, traces and… by @vvlrff in #3600
• fix(gemini): send the skip placeholder for unsigned current-turn function calls by @marklysze in #3603
• fix(a2ui): speak the stable A2UI v1.0 wire and re-vendor the spec by @vvlrff in #3618
• feat(mcp): let browser hosts connect, explain MCP Apps misuse, and ad… by @vvlrff in #3613
• fix(openai): accept streamed tool calls without an index by @pcbeingused333 in #3611
• chore(deps): bump provider SDKs in uv.lock by @vvlrff in #3621
• ci: install just and prek from PyPI instead of GitHub Releases by @vvlrff in #3622
• fix(mcp): read path-like MCP app documents as UTF-8 by @pcbeingused333 in #3608
• fix(tools): preserve wildcard prefixes in recursive file searches by @Ytq1115 in #3529
• fix(eval): make ag2.eval strict-clean and check it with mypy by @vvlrff in #3624
• fix(types): check events, knowledge, streams, middleware, observers and files with strict mypy by @vvlrff in #3625
• fix(types): check ag2.response and the core root modules with strict mypy by @vvlrff in #3626
• fix(types): check ag2/agent.py with strict mypy by @vvlrff in #3627
• fix(types): check ag2/extensions with strict mypy by @vvlrff in #3628
• chore(deps): bump astral-sh/setup-uv from 10.3.2 to 10.3.3 in the github-actions group by @dependabot[bot] in #3631
• docs: align custom expectation evaluator with the Hub protocol by @shuaiyafei in #3633
• docs: update Arize agent evaluation resource URL by @arizedatngo in #3630
• feat(extensions): add E2B sandbox integration by @max-sudolabs in #3320
• fix(ag-ui): send the reply a model call leaves only in its ModelResponse by @coder-abu in #3635
• feat(extensions): add boxd sandbox integration by @MichielMAnalytics in #3606
• chore: bump version to 1.3.1 by @vvlrff in #3636
• feat(mcp): serve upstream ext-apps App documents as written; mark MCP-UI legacy by @vvlrff in #3614

New Contributors

• @Ytq1115 made their first contribution in #3529
• @max-sudolabs made their first contribution in #3320
• @coder-abu made their first contribution in #3635
• @MichielMAnalytics made their first contribution in #3606

Full Changelog: v1.3.0...v.1.3.1