Agent Zero v2.0

Highlights

• Native Responses API transport — Agent turns now prefer the OpenAI Responses API when available, with automatic chat-completions fallback for providers that don't support it. Responses metadata is persisted in history and agent state for provider-state continuation, replay, and function-call execution. Prompt caching is normalized per provider (OpenAI/Azure, Anthropic, Gemini, Bedrock, OpenRouter).

• Parallel tool calls — A new parallel tool-call runtime lets the agent batch independent tool calls concurrently as visible child chats. Parallel await timeouts are non-destructive so jobs can be re-awaited, and nested subordinate chats work normally inside parallel workers.

• Revamped MCP server configuration — MCP servers now support project-scoped configuration with global/project merge semantics, a richer settings UI with list-first layout, server and tool search, per-tool enable/disable toggles, a dedicated scanner modal, and chat composer access. MCP runtime handling is hardened for dotted tool names, timeouts, status accuracy, lock scope, and cleanup of wedged transports.

• Time Travel workspace selector — The Time Travel plugin now supports selectable workspaces (workdir/project), with a workspace picker in the UI and new API endpoints.

• LiteLLM global kwargs — Normalized global LiteLLM configuration handling ensures framework defaults, configured globals, and per-call overrides merge correctly. Global kwargs are kept as per-call options rather than frozen into provider defaults.

• Shared canvas surfaces — Files is promoted to the universal canvas rail alongside Browser, Editor, and Desktop. All surfaces use a shared draggable/resizable floating modal with Focus mode. Surface windows persist across page refresh, and the right-canvas rail is visible on mobile (routing to modals on narrow screens).

• Welcome screen refresh — Reworked welcome screen with a shared new-chat composer, quick actions, discovery cards, OAuth accounts, and a system resources panel. The welcome screen stays free of docked canvas surfaces.

• Skills improvements — A new Settings > Skills scanner section runs security checks on skill zips before import. Loaded skills now persist through chat compaction by storing skill IDs as chat-wide context data and reattaching bodies post-compaction. Browser automation skills gain trigger frontmatter for better discovery. The a0-development skill is refactored into focused reference files.

• Protocol prompt area — A new prompt region ("protocol") is injected before message history, with per-turn temporary and persistent slots. Project instructions and active/loaded skills are now delivered through protocol rather than extras.

• A2A streaming — Agent Zero now advertises streaming capability on A2A endpoints by default.

New local provider support

• oMLX — First-class local provider for Apple Silicon with Docker-friendly defaults and paged SSD cache documentation.
• llama.cpp / llama-server — Registered as chat and embedding providers with host.docker.internal defaults.
• vLLM — Registered as chat and embedding providers; empty tools arrays are stripped for strict OpenAI-compatible servers.
• LM Studio & Ollama — Shipped Docker-friendly defaults so containerized Agent Zero reaches host-running services without manual api_base edits.
• Tiny Local tool-use profile — Bundled agent profile for small local models.

Telegram integration

• Major improvements to native Telegram integration UX and streaming.
• Telegram tool lifecycle hooks are now safe no-ops for non-Telegram contexts.

Security & dependency updates

• LiteLLM upgraded to 1.88.1 (above CVE-2026-42271 patched floor); OpenAI SDK pinned to 1.88.1.
• Starlette pinned to patched 1.0.1 release for Host-header advisory.
• requirements2.txt folded back into a single requirements.txt.

Settings & UI improvements

• Settings sidebar now has search with filtered results; sidebar accordion expansion is decoupled from active section; content pane scrolls independently. Backup & Restore promoted to its own section.
• Context window size control moved above the Advanced disclosure for main and utility model slots.
• Stale preset kwargs are cleared on model switches so unsupported parameters don't leak across providers.
• Rubik is now the default WebUI font, including native browser controls and the login page.
• Shared .text-button CSS primitive restored so controls render correctly without chat composer dependency.
• WebSocket heartbeat interval and timeout raised with environment overrides to prevent disconnects during long operations.
• Editor generalized from Markdown-only to exact-text .md and .txt documents with open, save, Save As, and rename support. "Open in Editor" is a visible File Browser row action.
• Google Cloud Gemini OAuth provider renamed for clarity.
• What's New showcase modal with version gating, permanent opt-out, and bundled media for new features.
• Project extension data hooks let plugins contribute named project settings sections without coupling to core helpers.

Bug fixes

• Pager CPU spin — Disabled pagers (PAGER=cat, GIT_PAGER=cat) in non-interactive code execution shells so git diff/git log no longer block forever at 100% CPU. Fixes #1697.
• MCP settings apply NameError — Fixed undefined variable when applying global MCP servers.
• Compaction backups with malformed Unicode — Surrogate-containing chat content no longer breaks JSON/transcript backup writes.
• Parallel worker chat log orphans — Ephemeral parallel worker chat folders are cleaned up; directories without chat.json are ignored at startup.
• Subordinate agent profiles — Profile arguments are validated against available profiles; profiles persist in saved chats and survive profile switches and restarts.
• Sentence-transformer embedding aliases — Stale provider/model aliases are patched at the config boundary so memory embeddings use the intended local HuggingFace provider.
• Project skills folder initialization — .a0proj/skills is created during project setup and repaired when loading existing projects.
• Connector file chunking — Chunked connector_file_op_result payloads are buffered and reassembled before resolution.
• Responses API 404 fallback — Wrapped NotFoundError responses from providers without the Responses API now correctly fall back to chat completions.
• Document query image routing — Images, screenshots, and diagrams prefer vision_load when vision tools are available.
• Remote tool prompt gating — Remote tool prompts are hidden unless the connected CLI advertises the matching capability.
• Mobile fixes — Welcome composer wrapping, project action layout, compaction modal spacing, modal layering over canvas rail, and composer alignment polished for narrow viewports.
• Unsecured connection settings link — The "Configure credentials" banner now opens Settings at the correct section.
• Welcome composer idle placeholder overlap — Ghost placeholder suppressed until a chat is selected.
• Xpra amd64 Docker fallback — Minimal fallback package set now works on amd64 when the preferred Xpra repository is unsatisfiable.
• pytest capture import — Guard added so pytest capture objects work with the code execution TTY helper.

Infrastructure & Docker

• Self-update manager synced at startup when the installed copy is stale.
• Uploads directory created during Docker runtime initialization before services start.
• Docker open file limit raised at startup with explicit compose nofile example.
• Comprehensive file-level DOX documentation added across the repository with matching AGENTS.md indexes.

Documentation

• New Launcher guide with runtime setup, Installs, Instances, and screenshot notes.
• v1.20-to-v2.0 image upgrade path documented with backup/restore guidance.
• README install section reorganized into Launcher, A0 Install, and direct Docker paths with A0 Launcher v0.9 download matrix.