v0.12.02026-06 · In preview
Parallel Sessions & Design Mode redesign
Two flagship features, in preview on the active branch (not yet released): fan one prompt out to Claude, Codex, and Gemini in parallel with a synthesizing orchestrator, plus a Figma-style, AI-first overhaul of Design Mode.
Parallel Sessions
- Fan-out multi-CLI: the same task launched on Claude, Codex, and Gemini simultaneously
- One-click launch from a "⇉ Parallel" pill in the composer, or an @oxima-parallel-* mention to preselect providers
- Autonomous, headless runs that never interrupt you mid-run
- Live council card streaming per-agent progress in real time
- Orchestrator benchmarks every result (0–10) and synthesizes the strongest ideas into a single plan
- Plan review gate (approve / modify / decline) before anything executes; configurable orchestrator model
Layers & navigation
- Layers Panel: Figma-style component tree with two-way selection (click tree ↔ click canvas)
- DOM ⇄ File view: switch between the live hierarchy and a group-by-source-file view, which also surfaces Assets (images/fonts) and Dependencies inline
- Per-instance targeting: a single `.map()` render is identified exactly — never "all of them"
Per-element inspector
- Edit text, background, text color, radius, shadow, border (width/style/color) and opacity — live preview before commit
- Add a class when the element has none, not just swap an existing one
- Live text editing with in-canvas preview across headings, paragraphs, spans, links, buttons, list items and more
- Every edit commits to the exact source line (snapshot-driven, never corrupts source)
Motion & AI-first edits
- Motion Picker: visual animation presets (on-load / scroll / hover / tap × easing × duration) compiled into a precise framer-motion / GSAP instruction
- 3-tier model: instant token tweaks stay direct; everything compositional (gradients, effects, animations) becomes a clean prompt the agent implements
- The bundled `oxima-design-engineer` skill now authors designs that are click-to-edit by default
Canvas & workspace
- Inspector as a bottom sheet that rises from behind the canvas dock
- Freely-draggable canvas dock (viewport presets, zoom, fit) that moves with the inspector as one unit
- Project Assets strip with drag-drop uploads to `public/`; git-backed per-turn snapshots with restore + apply-to-main
- Terminal improvements: multi-tab, search, font zoom, AI shortcuts
v0.11.02026-05
Agents, Telegram & Service Catalog
The cycle that turned OXIMA from a chat client into a connected workspace: domain-expert agents, a Telegram bridge, and one-click service + MCP catalogs.
Agent System
- @mention activation of domain experts — session-scoped and cumulative
- 9 bundled agents: @typescript, @react, @nextjs, @vue, @svelte, @tailwind, @supabase, @prisma, @tauri
- Per-agent knowledge sources with auto-discovery and AI-assisted rule updates
- Post-chat learning loop (`@compactor`) that extracts reusable rules from each session
- Export/import agents as portable `.oxima-agent` bundles
Telegram Bridge
- Run chat sessions from Telegram with one forum topic per session
- Voice-message transcription (Whisper/Gemini) and photo attachments
- Answer permission, plan-review and AskUserQuestion prompts via inline keyboards (first responder wins)
- `/model` to switch provider+model, `/prompts` to fire your prompt library, `/stop` to interrupt
Service & MCP Catalog
- One-click install + auth of 14 cloud services (Vercel, Netlify, Railway, Cloudflare, Firebase, Supabase, Sentry, GitHub, GitLab, Fly.io, Heroku, DigitalOcean, Stripe, PlanetScale)
- Curated third-party MCP server catalog (GitHub, GitLab, Slack, filesystem, memory) — plus add any server
- Lazy npx/uvx install, OAuth + env-token auth, atomic writes to Claude and Codex configs
- MCP tool inspector: handshake to list a server's tools before connecting
- Cross-CLI token sharing via `~/.oxima/secrets/<id>.env`
Providers & Sharing
- OpenRouter and more OpenAI-compatible cloud connectors
- Design Share: one-click public ephemeral preview link (`share.oxima.ai/p/<slug>`)
- Claude Opus 4.8 support
- New OXIMA theme — brings presets to 24
- AskUserQuestion session-routing fix + trusted-folder handling
v0.10.02026-04-18
Design Mode & Local Models
The two flagship features of this cycle: build interactive webapps inside OXIMA, and run any open-source model locally or via OpenAI-compatible cloud.
Design Mode
- New `kind: 'design'` chat session with a Vite + React + Tailwind sandbox scaffolded per session
- Live render in an integrated iframe — preview always up to date
- Edit via devtools-style comment: click an element → breadcrumb, CSS selector, computed style → write feedback → the agent applies the change
- Pin overlay for multiple comments, batch edits in a single request
- Auto-paused dev server during edits to avoid Vite flicker/errors
- Automatic git checkpoints per turn with branching for variant exploration
- Markdown blueprint export (mirror prompt) for handoff to a real-project agent
- Bundled `oxima-design-engineer` skill that guides the agent on stack choices per domain
Open Source / Local Models
- OpenAI-compatible HTTP client with SSE streaming and tool-calling loop (read_file, write_file, edit_file, run_command, glob_search, grep_search, list_directory)
- Reasoning support (`delta.reasoning_content`) for thinking-mode models like Qwen 3 and DeepSeek R1
- Ollama integration: detect binary, list installed models, pull with progress, delete, open Terminal for download
- Curated catalog: Qwen 3 (0.6B → 32B), Gemma 4 (e2b → 31b), DeepSeek R1 (1.5b → 32b), GLM-4 9b
- Custom endpoint modal with presets (Ollama, LM Studio, vLLM, Custom URL)
- OpenAI-compatible cloud connectors: Groq, Together.ai, Fireworks.ai
Polish
- Universal preview adds JSX rendering and Feedback Overlay annotations (8 formats total)
- Process-group kill for child processes (npx, vite, esbuild) — no more orphans on crash
- Plan Mode fixes and permissions hardening
v0.9.02026-04
Private Beta
The release that made OXIMA ready for the first external beta testers. Theming, skills, automations, and the workflow polish needed for daily use.
Theming
- 22 preset themes (Forge Dark, Monokai, Solarized, Nord, Dracula, Tokyo Night, and more)
- Custom theme builder with base + accent color picker
- Sidebar transparency 0-100% with backdrop blur (glass effect)
- Zero React re-render — everything via CSS cascade
Agent Skills
- Auto-discovery from `.claude/skills/`, `.agents/skills/`, `.gemini/skills/`, `~/.claude/skills/`
- Open standard with YAML frontmatter (name, description, triggers, resources)
- Skill Creator UI with automatic frontmatter generation
- Skill Installer with cross-CLI sync
- Skills Browser
Automations
- Scheduler V3 with intervals (every N minutes/hours) or specific 24h times
- Weekday filter, one-shot or recurring
- Target session (reuse existing or create a new one per run)
- 18 ready-made templates with full execution history
- Dual-layer concurrency guard + crash recovery on restart
Scaffolding & Templates
- 34 templates across 12 frameworks (Next.js, React, Vue, Nuxt, SvelteKit, Angular, Astro, Expo, React Native, Remix, Solid, Flutter)
- Variants per package manager (npm/pnpm/yarn/bun) where applicable
- Categorized templates: Image Generation, MCP Setup, LLM Models
Git & Workflow
- Branch selector in the top bar
- Commit and push from the app, with auto upstream (`-u origin HEAD`) on first push
- Diff viewer with auto-refresh every 10s, staged/unstaged separation
- Popout chat for multi-monitor workflows
- Context Panel: Progress Card stepper + Artifacts Card with AI-touched files
v0.8.02026-03
Multi-Provider & Chat Hardening
The release that unlocked parallel multi-provider orchestration and made chat sessions resilient under real-world load.
Multi-Provider Orchestration
- Codex CLI integration alongside Claude Code
- Gemini CLI with stream-json, tool events, auto-fallback Pro → Flash on 429
- Per-session provider and model, persisted on `ChatSession`
- Provider/model restored on project or session switch
- Concurrent chat streams with backend registry (`ACTIVE_CHAT_STREAMS`) and per-stream `cancel_flag`
- Configurable extended thinking budget (low 2K / medium 10K / high 50K / max 128K)
Resilience
- Auto-retry with progressive backoff (5s/10s/15s) on transient errors (tool_use concurrency, rate_limit, 5xx)
- Post-stream retry: when CLI exits after streaming partial content, `reset:true` event clears the partial message and respawns
- AskUserQuestion dialog: the AI can ask multi-option or free-text questions
Chat UX
- Plan Mode with dedicated review dialog ("Execute Plan" / "Discard")
- Speech-to-text via Whisper — record and transcribe audio in the prompt
- IDE context toggle: inject `git status` + `git diff --stat` into the prompt
v0.7.02026-02
Foundation
The first stable release with the architectural core: native desktop app, Claude Code integration, hooks, permissions, file explorer, terminal — the foundation everything else builds on.
Architecture
- Tauri v2 + React 19 + Rust: native desktop app (not Electron)
- Claude Code CLI integration with stream-json streaming
- Atomic state persistence (temp file + rename) for zero corruption on crash
- Multi-project workspace with isolated state, chat, configuration
- Auto-setup of hooks + MCP for every new project
Permissions & Hooks
- Hook server (Axum) on ports 9876-9890 for PreToolUse / PostToolUse / PermissionRequest
- Hook-based permissions system with two modes (Active dialog UI, Spectator auto-approve)
- Smart timeouts: PreToolUse 15s auto-ALLOW / PermissionRequest 60s auto-DENY
- Four Interaction Modes: Ask, Auto Edit, Full Auto, Plan
Workspace Tools
- Project File Explorer: hierarchical tree with 36+ language icons, search, attach-to-chat
- Built-in Terminal: native PTY (xterm.js + portable-pty), zsh/bash login shells, real-time streaming, dynamic resize
- Tool call groups in chat: collapsible blocks with icons per tool type
MCP & Capacity
- Native MCP server (stdio) with 8 tools: forge_get_tasks, forge_report_progress, forge_mark_complete, forge_signal_done, forge_create_task, forge_create_plan, forge_remember, forge_recall (a 9th, forge_get_current_task, was added later — current total is 9)
- Capacity dashboard with quota tracking and budget policy (Pro 1x / Max5x / Max20x)