First public release: chat, specialists, sandbox, editor

0.1.7 is the first formal release of the AI platform. Chat, custom specialists, an embedded code editor, a permissioned tool sandbox, real-time workflows, and the iOS and Android apps all ship together as a single product. Billing, OpenRouter and local-model providers, and account-aware sign-in round out the foundation.

Chat

Chat is the home surface. Start conversations, mention specialists and people, drop attachments and tagged files, and pick up where you left off — your draft message persists per conversation across reloads. Messages support translations, voice input, and live progress updates as a specialist works. Reactions, presence, and unread state stay in sync across the workspace, and a ribbon above the input lets you see when streaming or auto-replies are paused.

Chat now supports rich-block messages, live tool output, and an embedded browser tab inside the conversation for previewing links without leaving chat. Mention search, specialist mention ranking, and embedded browser positioning have all been hardened for this release.

Specialists

Specialists are the agent layer. Create custom specialists from a guided wizard, pick the model behind them (OpenRouter, Hugging Face, or a local model with optional GPU acceleration), and choose which tools they can use through a unified permission system. A multi-format definition editor, draft renaming, and a dedicated chat channel for the creation flow make iterating on a specialist feel like authoring a coworker rather than configuring a backend.

Specialists can read and write project files, drive the embedded browser, run interactive shell commands inside the sandbox, and use code-aware tools like rename, go-to-definition, and find references. They can also be assigned to tasks, post live progress updates, and emit rich chat blocks. Project containers tie a specialist’s work to a repository so a specialist actually has somewhere to go to work.

The Figma specialist suite, the Google Workspace specialist suite, and a benchmark IQ scoring loop are all live, and the specialist builder ships with a test mode so you can validate a specialist before publishing it.

Editor

A VSCode-based editor is embedded inside the app. It opens files reliably, supports terminals, ships language-server support, syncs its theme with the app, and exposes an API so the rest of the app — and specialists — can drive it. Code panel, browser panel, and chat panel are independently collapsible, and “Open in IDE” buttons on code blocks pull you straight from a chat message into the editor.

The file layer underneath the editor watches files, syncs writes back to the local filesystem, supports search, and applies a sandbox profile that blocks reads and writes a specialist hasn’t been granted.

Sandbox

The sandbox is where specialist tool calls actually run. It supports stdin and environment variables, an interactive command path, a package manager for installing tooling on the fly, and a Windows implementation alongside macOS and Linux. The agent skills system (agentskills.io) plugs in here, the AI Bundler runtime is integrated for remote-source profiles, and tool permission requests now show inline cards with the tool name and the reason a specialist needs the access.

A bundled code knowledge graph gives specialists structural understanding of a repository — symbols, references, and call sites — and an autoresearch loop lets specialists do iterative web research as a tool.

Workflows

Workflows are a real-time graph editor for orchestrating multi-step pipelines. Drop in nodes for git history analysis, git blame, file modifications, JavaScript execution, and dynamic file pickers; chain them together; and run them with async node execution. Workflows can invoke custom specialists, create tasks, and attach to channels and specialists. An async PRD QA loop drives interactive QA passes.

Settings, billing, and providers

Settings is a fully redesigned dialog with mobile-responsive drawer navigation, UI scale controls, Connections for OpenRouter and Hugging Face model management, and appearance controls. Local models, when enabled, contribute their own token usage so you see the same accounting whether you are running cloud or local. Billing is live with production credentials and a release-channel switcher in the app, so you can move between stable, beta, and nightly without reinstalling.

Auth, projects, and tasks

Sign-in supports multiple accounts, organization management, an anonymous pre-auth state, QR-code login between desktop and mobile, and session-expiry recovery that returns you to a clean workspace on resume. The login screen and the Auth0 sign-in template have been redesigned together so the handoff in and out of the OAuth window feels native.

Projects are containers tying repositories, tasks, and conversations together, with smooth loading across the project list, project detail, and task views. Tasks can be assigned to specialists, automatically open a dedicated channel, and carry attachments. A pull-request QA flow uses isolated worktrees so QA runs do not stomp the working copy.

Mobile, platform, and performance

iOS and Android apps ship in this release with mobile OAuth login. On the desktop side, the app picks up updates automatically, hides to the macOS dock and the Windows and Linux tray, supports drag from empty sidebar space, and applies macOS window corner radius and proper transparency on Windows and macOS. Startup is faster, and the embedded browser uses native Chromium on macOS and Windows.

Polish and fixes

Hundreds of fixes land alongside the headline features. Specialist mention ordering, embedded browser positioning, and the OAuth callback window are all stabilized. Permission grants reuse “always allow” choices for the same workspace and specialist. The code knowledge graph resumes safely under thermal pressure. Real-time sync, sign-in connections, post-create calls in the Google Workspace suite, and theme changes are all hardened. Workspace creation and project dialogs no longer over-fetch, and the chat panel is always collapsible.