← Back to changelog
v2.0

Parallel Tools, Responses API, & MCP Redesign

June 23, 2026

v2.0 is a major release focused on stronger tool orchestration, native model transport, richer project configuration, safer skill workflows, and a more polished WebUI for daily agent work.

⚡ Parallel Agent Work

  • Inspectable parallel tool calls — Independent work can run concurrently while preserving visible child-chat steps and follow-up awaits.
  • Safer parallel batching — Parallel-call guidance now distinguishes independent tool batches from unsupported recursive or background cases.
  • Follow-up await handling fixed — Await timeouts no longer turn incomplete parallel work into destructive failures.
  • Parallel worker chat logs fixed — Worker chat history is kept attached to the parent workflow instead of creating orphaned records.

🔌 MCP, Skills, & Project Context

  • Redesigned MCP configuration — Global Settings and Projects now share a cleaner MCP setup flow with command and Remote URL transports, search, raw JSON mode, and safer removal behavior.
  • MCP timeout hardening — Wedged MCP transports and cleanup paths are less likely to hang the agent loop.
  • Skills Scanner flow — Skills can be checked before import, with safer package extraction and improved Skills settings search.
  • Protocol prompt area — Project instructions and active skill context now live before history, keeping durable task context clearer.
  • Loaded skills survive compaction — Active skills are reattached after compaction while avoiding unnecessary prompt bloat.

🧠 Models, Providers, & Transport

  • Native Responses API transport — Responses API support now includes native function-tool handling, stored response metadata, prompt-cache awareness, and fallback behavior for providers without /v1/responses.
  • Predictable LiteLLM globals — Framework defaults, configured global kwargs, and per-call overrides merge more consistently.
  • Cleaner preset switching — Provider-specific kwargs are cleared when switching model presets so stale values do not leak into the next provider.
  • Local provider presets — LM Studio, Ollama, oMLX, llama.cpp, and vLLM have first-class local provider support with Docker-friendly defaults and onboarding metadata.
  • Context window visibility — Model settings expose context window size more directly.

🖥️ WebUI & Workflow

  • Refreshed welcome and composer surfaces — The welcome screen, composer, setup areas, and placeholder behavior have been polished across desktop and mobile.
  • Shared canvas surfaces — Files now participate in the shared canvas/sidebar surface flow alongside Browser, Desktop, and Editor.
  • Browser live view restored — Browser defaults back to live CDP viewing with snapshot fallback.
  • Editor text workflows improved — Exact-text Markdown and text files can be opened, saved, saved as new files, and renamed more reliably.
  • Time Travel workspace selection — Time Travel can target selectable workspaces.
  • What's New showcase — A built-in version-gated showcase highlights major new capabilities such as parallel work, MCP configuration, and Skills Scanner.

📱 Mobile, Connectors, & Telegram

  • Mobile layout polish — Composer alignment, modal spacing, Browser toolbar controls, Projects actions, canvas rail behavior, and compaction dialogs received focused fixes.
  • Connector file results reassembled — Chunked connector file results are reassembled before use.
  • Remote tool prompts gated by metadata — Remote prompts now respect connector metadata before exposing host tool instructions.
  • Telegram integration improved — Command scoping, streaming behavior, native media delivery, session picking, long-run status updates, and friendly errors are all stronger in this release.
  • Document/image routing clarified — Screenshot, scan, photo, diagram, chart, and image-heavy workflows are steered to vision tools while document-heavy workflows stay with document_query.

🛠️ Reliability, Packaging, & Docs

  • WebSocket heartbeat grace extended — Long prompt and context work has more room before heartbeat timeouts interrupt the session.
  • Docker startup hardening — Upload directories are created on startup, stale self-update manager scripts are synchronized, and amd64 Docker builds can fall back to a minimal Xpra package set when preferred packages are unavailable.
  • Compaction backup resilience — Malformed Unicode no longer breaks compaction backups.
  • Security and compatibility pins — LiteLLM, OpenAI SDK, and Starlette pins were refreshed for current compatibility and security fixes.
  • Documentation contracts expanded — File-level DOX coverage and AGENTS indexes were refreshed across runtime, plugin, API, WebUI, helper, tool, and skill areas.
  • a0-development references grounded — Development guidance was split into more focused reference files.