This is the speed release, with some serious resilience work underneath. We halved the time between your prompt and the first token in Agent Zero. In the Launcher, startup time is also cut in half, and saving settings and Instance operations are about 30x faster. Context Doctor: A brilliant community contribution that brings a new recovery layer that repairs malformed tool-call JSON and recovers from empty, repeated, or misformatted model responses. If you run small local models or providers that occasionally misbehave, your runs stop derailing. Context-window indicator: You can now see live token usage, remaining context space, provider-reported costs, cache hits, and generation updates while the agent works. No more guessing what a run costs or when compaction will kick in. Vision sidecar: Another community contribution. Configure a separate vision model alongside your main model. Perfect when your favorite text model has no vision support, or when you want a cheaper model handling images. Also in Core: - The WebUI is faster and steadier during high-rate streaming with progressive Markdown and LaTeX rendering. - Transports now default to Chat Completions, which makes life easier for plugin developers bringing in new providers. - The interactive Browser gets configurable keyboard layouts with better macOS Option and AltGr handling. CLI: - WebSocket recovery keeps running through long outages with safe ownership transfer during reconnects, host changes, and shutdown. - Authorized Chrome/CDP sessions are more stable with persistent browser identities. - Now you get terminal-native ready notifications when TUI or headless runs complete. Launcher: - Startup, Versions, Settings, and Instance operations reuse current data, so everything feels instant while explicit refreshes stay authoritative. - There is a brand new Advanced section for power users: a file-first workspace for Dockerfiles and Compose files with fixed build, run, and teardown actions. - Running local Instances get a native Restart that waits for the Agent Zero UI to be ready.

Three main dishes this time: a Browser that works on the real web, @ tagging for everything, and ACP support in the CLI. The Browser got stealthy: The internal Browser now handles bot detection properly. Sites that flagged Agent Zero as a bot and were unusable in previous versions now just work, including Gmail and your social accounts. You sign in yourself through an interactive authenticated viewport, then the agent works inside your session. The Browser runtime is shared across chats: sign in once, every chat can use it, tabs restore automatically, and cleanup is safer. On top of that: native resizing, DOM hover support, multi-page annotations, clearer loading feedback, and voice controls. The attached video shows it browsing Gmail. Tag anything with @: Type @ in the composer and get scoped autocomplete for agent profiles, files, folders, skills, and MCP servers. No more typing full paths or remembering exact names. It works in the Web UI and in the CLI, where completion also covers bounded local or container workspace paths. Composer drafts now persist per chat too, so a reload never eats your half-written message. The CLI speaks ACP: Run 'a0 acp' and the CLI starts an Agent Client Protocol stdio server backed by the Agent Zero connector, with session lifecycle and host-tool support. In practice, external editors that support ACP can now host Agent Zero directly: your editor becomes the frontend, Agent Zero does the work. Also in this release: Unified subordinate agent lifecycle across direct and parallel calls with resumable failed children, streamed tool calls preserved when final response envelopes omit them, plugin reload prompts scoped to global changes only, authenticated extension routes with path-traversal protection, and backup restore that no longer affects the active process settings. How to update CLI: open a terminal and type 'a0 update' Web UI: open Settings -> Check for updates -> Review update New here? Install guide: https://www.agent-zero.ai/p/docs/installation

This one is about stability and making the first five minutes better. What changed for new users: The first-run setup was reworked around an in-chat model gate. You can type your first message before configuring anything — once your model is set up, Agent Zero sends it automatically. Provider selection (cloud API keys, connected accounts, local models) is now one unified flow, and Main/Utility model selection is explicit instead of silently defaulting after OAuth. What changed under the hood: Big improvements to LiteLLM and Responses API handling. If a provider or proxy rejects or mishandles /v1/responses, Agent Zero now falls back to Chat Completions cleanly — and streamed tool calls survive the fallback as structured results, so tools keep working. Parallel tool extraction and numeric argument handling were also fixed. Defaults: The bundled main model, Balance preset, and onboarding OpenRouter default are now Claude Sonnet 5. CLI v2.2: The CLI can now discover local Agent Zero instances through Colima and other Docker socket endpoints (DOCKER_HOST, Docker contexts, known runtime paths), shows an install hint when nothing is found, refreshes chat tab names in the TUI, and includes computer-use fixes for macOS permissions and Ubuntu helper launch. Full changelogs at github.com/agent0ai/agent-zero How to update CLI: open a new terminal and type 'a0 update' Web UI: open Settings -> Check for updates > Review update

This is our first major release after the UI redesign of the version 1 line, and it's really three things shipping together. 1. A0 Launcher (our new product) A desktop app for Windows, macOS and Linux. It manages your local and remote Agent Zero instances from one place, and, for anyone just starting, it onboards new users with an installer that sets up the container runtime automatically. You don't need to understand Docker to get running. Visit https://www.agent-zero.ai and look right above the install scripts. There is a download link for your platform and a link to a Launcher first run guide. - Mac Intel - https://github.com/agent0ai/a0-launcher/releases/download/v0.9/a0-launcher-0.9-macos-x64.dmg - Linux x86 - https://github.com/agent0ai/a0-launcher/releases/download/v0.9/a0-launcher-0.9-linux- - Windows x86 - https://github.com/agent0ai/a0-launcher/releases/download/v0.9/a0-launcher-0.9-windows-x64.exe ARM64 - Mac Apple Silicon - https://github.com/agent0ai/a0-launcher/releases/download/v0.9/a0-launcher-0.9-macos-arm64.dmg - Linux ARM64 - https://github.com/agent0ai/a0-launcher/releases/download/v0.9/a0-launcher-0.9-linux-arm64.AppImage - Windows ARM64](https://github.com/agent0ai/a0-launcher/releases/download/v0.9/a0-launcher-0.9-windows-arm64.exe) | 2. Agent Zero Core v2.0 The engine itself: native Responses API transport (with chat-completions fallback), inspectable parallel tool calls, project-scoped MCP configuration with search and scanning, a Skills Scanner for checking packages before import, and a refreshed WebUI across welcome, composer, settings, mobile and the shared canvas. Full changelog at https://github.com/agent0ai/agent-zero/releases

With v1.19, we are aligning the version numbers of the A0 CLI and the Agent Zero Web UI. Both are now on the same version, which should make updates and support much cleaner going forward. But the more important part is what this release makes possible. 1. DOX: project context that agents can actually follow We are introducing DOX, our open-source layered documentation framework. Think AGENTS.md, but not trapped in one flat file. With DOX, each part of a project can describe itself. The agent gets the right context where it is working, instead of trying to hold the whole repo in one giant instruction blob. This should make agents much more reliable in real projects, especially larger codebases. 2. Stronger Computer Use Computer Use is also stronger across Linux, Windows, and macOS. The big direction here is simple: agents need to work with real computers, not just chat interfaces. v1.19 moves us further in that direction, including background computer use. 3. Faster document_query with Liteparse We also upgraded document_query with a Liteparse-powered engine for much faster PDF parsing and better document Q&A. This is the kind of feature that becomes very useful once you start giving agents real research tasks: PDFs, docs, reports, claims, contracts, benchmarks, and anything else where the answer needs evidence. 4. Plugin toggler and chat tabs in the CLI Now you can switch context more easily in the CLI thanks to the tabs in the upper area. And you can also enable/disable plugins with the /plugins slash command. Thank you to our stakers and users. Every week, stakers used 6 billion tokens through the Agent Zero LLM API, plus more than 2 billion daily tokens on OpenRouter. That is real usage, and it helps us keep pushing the product forward. If you try v1.19, please share: - what Computer Use works well on your OS - where document_query feels faster or still struggles - any workflow that breaks The best releases are the ones the community stress-tests with real-world use cases.