Ava Supernova – open-source coding agent for open-weight LLMs

Ava Supernova, an open-source AI coding agent, launches with support for open-weight LLMs, offering 60+ tools, 7 modes, and 24 specialist personas. The agent is free to start, local-first, and requires no account or credit card, aiming to democratize agentic coding by running on frontier open-weight models that match closed models at lower cost.

The open-source AI coding agent that works for everyone — not just the people who can afford it. 60+ tools · 7 modes · 24 specialist personas · frontier open-weight models from every major lab · a Creative Studio that builds real assets into your project. Free to start, no account, no credit card. What Ava Is what-ava-is · Surfaces one-ava-four-surfaces · Models models--fleets · Getting Started getting-started · Tools tools · Privacy privacy · Contributing contributing Agentic coding shouldn't be a luxury. Ava Supernova is a full AI coding agent — she reads your codebase, plans the work, executes it task by task, and remembers what she learns so the next time gets easier. She runs on the frontier of open-weight models, which now match or beat closed frontier models on agentic coding at a fraction of the cost. She's open-source, local-first , and free to start: No account, no credit card, no trial. Install, open, go. Bring your own API key and run fully local — or sign up for 300 free credits a month and start with zero setup. Your data stays on your machine. Memory, history, journal, tasks — all local by default. Cloud sync is opt-in. No gatekeeping. Every capability — including the security toolkit and the teaching engine — is available to everyone. The free tier is a floor, not a free trial. Ask Ava what can you do? and she'll read her own source code and tell you herself. The same agent engine @ava/core powers every surface — same tools, same personas, same memory. You choose where you work: | Surface | What it is | Best for | |---|---|---| CLI @ava/cli | A terminal REPL with the full agent loop. | Living in the terminal, scripting, servers, SSH. | VS Code Extension | Ava inside the editor you already use — chat, agent, command-centre dashboard. | Adding Ava to your existing VS Code workflow. | IDE | A standalone desktop app Tauri + Rust with Ava native and a local engine. Owns desktop automation. | A dedicated AI-native workspace; controlling the whole machine. | Companion | A mobile/web app — chat, tasks, journal, memory on the go. | Ava away from your desk. | She remembers you. A layered memory system learns your style, conventions, decisions, and the bugs you've fixed before — curated by a dedicated Memory Agent so context stays lean and recall stays sharp. Local-first, opt-in semantic recall. She thinks before she builds. 24 specialist personas across 7 modes plan, verify, and challenge before executing. Simple questions skip orchestration entirely — zero overhead. She picks the right fleet for the work. You pick the job; the right multi-model fleet falls out of it see Models & Fleets models--fleets . She teaches for free. Switch to Teach mode, say "teach me Rust," and she builds a personalised curriculum and teaches it lesson by lesson — fact-checked against the live web. Same credits as every other mode. She explains herself. Ask "how does your memory work?" and she reads her own implementation and walks you through it. She's yours to design. Name, pronouns, tone, energy, communication style — same brain, your personality. Full UI in 20 languages. Modes aren't tool restrictions — they're states of mind. The toolkit and the persona team change to match the work. | Mode | Prefix | Mindset | Use it for | |---|---|---|---| Code | | Builder | Writing, editing, shipping code with the full toolkit | Plan | :: | Architect | Read-only analysis, feature design, strategic proposals | Chat | .. | Friend | Personal conversation, journal, memory, news, weather | Teach | ?? | Tutor | Personalised learning paths and Socratic delivery | Security | | Auditor | OWASP-aligned scanning with verified findings | Brainstorm | | Ideator | Research, generate, challenge, and refine ideas | Write | << | Author | Drafting real documents — reports, proposals, letters, articles — exported as branded Word & PDF | Switch with the dropdown or Ctrl+Shift+1 through Ctrl+Shift+7 . When a task is complex, Ava's internal team activates. Each persona gets its own clean context and scoped tools the Challenger reads but never writes . Simple tasks go direct. Code: Scout → Architect → Verifier → Sequencer → Challenger → Builder Plan: Researcher → Architect → Challenger Teach: Curriculum Architect → Content Writer → Fact Checker → Quiz Master → Tutor Security: Recon → Scanner → CVE Researcher → Verifier → Reporter Brainstorm: Explorer → Researcher → Ideator → Challenger → Refiner Chat and Write run without personas — just Ava. You don't pick a model — you pick the work, and the right fleet falls out of it. Two orchestrated fleets ship today; a third is rolling out. | Fleet | Status | Coordinator | Best for | |---|---|---|---| Maestro | Ships now | Qwen 3.7 Plus | Daily work, predictable cost — a tier-differentiated Qwen ensemble | Aurora | Ships now | Mistral Medium 3.5 | GDPR-strict / sovereign EU work — Mistral-only, open weights, never leaves EU infrastructure | Supernova | BYOK now · managed preview | DeepSeek V4 Pro | Heavy multi-step work — a polyglot fleet picking the best model per role | Supernova runs today with your own DeepSeek + Qwen keys; the managed platform-hosted version is in preview. Maestro unlocks with a Qwen key, Aurora with a Mistral key. Bring your own key for any provider, or use Ava-managed models on a plan. Every plan gets every managed model — tiers differ only on monthly credit allowance, never on which models you can use. | Provider | Models | |---|---| Ava Managed | Qwen 3.7 Plus, Qwen 3.5 Plus, Qwen 3.5 Flash, plus the managed Mistral Aurora coordinators | Qwen Alibaba | Qwen 3.7 Plus, Qwen 3.7 Max, Qwen 3.5 Flash — 1M context, vision, reasoning | DeepSeek | V4 Pro, V4 Flash — MIT open-weight, 1M context | Moonshot | Kimi K2.7 Code — 1T MoE / 32B active, 256K context | Zhipu | GLM-5.2 — MIT open-weight, 1M context | Mistral | Medium 3.5, Small 4, Large 3, Codestral, Devstral — EU, open weights, up to 262K context | MiniMax | M3, M2.7 — 1M context BYOK | Xiaomi | MiMo V2.5, V2.5-Pro — 1M context, native multimodal | Anthropic | Claude Opus 4.8, Sonnet 4.6, Haiku 4.5 | Custom / Local | Ollama, LM Studio, vLLM, or any standard API-format endpoint — the $0 path | Node.js https://nodejs.org v20.0.0 or later- An API key from any supported provider — or sign up for 300 free credits a month Install from npm npm install -g @ava/cli ava On first launch, Ava runs an interactive setup wizard: pick a provider, paste your key or skip and sign in for the free tier , and you're ready. Install from the VS Code Marketplace https://marketplace.visualstudio.com/items?itemName=augmentedvalueacceleration.ava-supernova — or search "Ava Supernova" in the Extensions panel. Click the Ava icon in the activity bar or Ctrl+Shift+A and start. Download the standalone desktop IDE from ava-supernova.com https://ava-supernova.com . A local Node.js engine runs @ava/core on your machine — the full agent without a browser or VS Code. git clone https://github.com/AugmentedValueAcceleration/ava-supernova.git cd ava-supernova pnpm install pnpm build 60+ user-facing built-in tools. Ava decides which to use, runs them, reads the results, and keeps reasoning. The IDE adds desktop- and browser-automation tools on top, gated to Desktop mode. | Category | Tools | |---|---| File ops | file read , file write , file edit , glob , grep , list directory , find symbol , project index , analyze architecture , self inspect , docs lookup , browse library | Shell | bash , test run , test generate , benchmark , debug logs | Git | git status , git diff , git commit , git create pr , rollback | Web | web search , http request , browser , weather , news , release notes , paper fetch full text | Media | generate image , generate video , remove background | Documents | document author , document manage , email draft , report generate , doc generate , present plan , apply plan , todo write , task manage , journal write , switch mode | Memory | memory save , memory recall , memory update , memory delete , curator | Learning | learning create , learning teach , learning progress | Security & Data | audit dependencies , database query , secret request , env write | System | get datetime , detect language , ask user , support request , propose tool | Ava generates real assets and wires them straight into your project — a logo into public/ , a hero clip into your landing page. Images & video run on Wan Alibaba's open-weight models , built for clean graphic-design output and short clips with synchronised audio.- Every asset is tracked in the Library — grid or list view, filter by type, preview, reuse across projects. Your data stays yours: API keys encrypted in your OS keychain VS Code SecretStorage . Never logged, never sent to Ava servers. Conversations & memory stored locally ~/.ava/ . Local-first by default. Credential safety — runtime detection blocks API keys, tokens, and private keys from being saved to memory; high-confidence key patterns are redacted from chat output before they hit the screen. Capability-based secrets — Ava receives opaque handles for vault secrets, not raw values, scoped to the current session. Prompt-injection resistance built in. No telemetry — we don't track your usage, code, or conversations. Cloud sync is opt-in , with delta tracking so you see exactly what's about to be pushed. Every plan gets every managed model. Tiers differ only on monthly credit allowance. | Tier | Price | Monthly credits | |---|---|---| Free | $0 | 300 | Pro | $19 | 5,000 | Ultra | $39 | 10,000 | Enterprise | $79 | 20,000 | Or run fully local with your own keys — the floor is free, forever. A pnpm monorepo with a shared core, plus submodules for the IDE, web platform, and companion app: packages/ ├── core/ @ava/core — shared agent engine agent loop, providers, tools, personas, memory ├── cli/ @ava/cli — terminal REPL ├── extension/ VS Code extension host + React webview + dashboard ├── ide/ Standalone IDE Tauri v2 + Rust + Node.js engine — submodule ├── mobile/ Companion app — submodule └── web/ Platform website — submodule Key design decisions: Shared core, zero duplication — CLI, extension, and IDE all run the same @ava/core engine, tools, and providers. Event-driven agent — the agent emits typed events every host subscribes to. Self-contained tools — each tool carries its own schema and handler; adding one is a single file plus one registry line. No provider SDKs — all provider communication uses native fetch . Permission matrix — risk levels safe , write , dangerous combine with your permission mode to decide what needs approval. Plans and user questions always require approval. Local-first — everything works offline; cloud sync is opt-in. Contributions are welcome — this is built for the community. - Fork the repository - Create a feature branch from development - Make your changes and run pnpm build - Open a pull request against development Apache License 2.0 /AugmentedValueAcceleration/ava-supernova/blob/development/LICENSE — Copyright 2026 Augmented Value Acceleration Agentic coding for everyone.