I remember how you build
I'm Ava — an open-source coding agent that learns your project, your taste, and your habits, and carries them across every session. Free forever, local-first, and yours on any model you choose — or fully offline on your own stack.
300 credits every month — enough to try every model, mode, and tool. No credit card. No strings.
Scout scanning 47 files across 6 directories
Architect analyzing patterns — Next.js 15, React 19, Tailwind 4
Verifier checking dependencies — 23 packages, 0 vulnerabilities
Challenger found 3 things worth addressing
Health & training, in the same agent
A 225-dish global recipe library. 116 canonical movement patterns. Three skill levels per dish. Ava-authored technique guides and demo imagery. Free, like everything else.
Shipped
- 225-dish recipe library across 60 regions
- Three skill levels per dish — beginner, intermediate, expert
- 116 canonical movement patterns — squat, hinge, push, pull, lunge, carry, core, plyometric, cardio, mobility, yoga, breathwork
- Ava-authored technique guides, beginner cues, common mistakes, plus hero and demo imagery
Coming next
- Personalised meal plans that know your pantry, taste history, and training schedule
- Adaptive workouts that respond to logged soreness, sleep, energy
- Mobility and breathwork sequences for recovery
- Companion intake — weekly logs feeding Ava’s memory so recommendations sharpen with you
Three fleets. One agent. Pick a strategy, Ava picks the model.
You talk. Ava classifies what you need and dispatches to the right specialist for each subtask. Three orchestration strategies, all available on every plan.
Maestro
Production-tunedTier-differentiated Qwen fleet. Qwen 3.6 Plus on the heavy slots, Qwen 3.5 Flash on chat and image-gen orchestration, Omni Plus on vision. Most cost-efficient, deepest production soak.
- Qwen 3.6 Plus coordinator
- 1M context, native MCP function calling
- Terminal-Bench leader on real agent loops
Supernova
Polyglot ensembleDeepSeek V4 Pro coordinates and dispatches: Qwen 3.6 Plus for Builder spawns, V4 Flash for chat + review, Omni Plus for vision. Frontier reasoning where it matters, flash-tier economics on the bulk.
- DeepSeek V4 Pro coordinator
- Best model for each role, not coordinator-only
- Cost-aware: heavy where needed, light everywhere else
Aurora
European AI stackMistral-only, three-tier. Large 3 coordinator + heavy specialists, Medium 3.5 for Builder + vision + long-form, Small 4 for chat + intent gate. Open weights end-to-end. Never leaves European infrastructure.
- Mistral Large 3 + Medium 3.5 + Small 4
- GDPR-strict, AI Act compliant, sovereign by design
- Apache-2.0 weights — no closed model in the loop
All three modes. Every major model. Or your own local stack.
BYOK works with all three orchestrated modes — same agent, same memory, same personas, just running on keys you own. Or skip orchestration entirely and drive a single model directly.
Managed providers
Kimi K2.6, Claude Opus 4.7 / Sonnet 4.6, DeepSeek V4, Mistral Large 3 / Medium 3.5, Zhipu GLM-5, Qwen 3.6 Plus, MiniMax M2.7. Add your key, run unmetered.
Local + offline
Ollama, LM Studio, vLLM, or any OpenAI-compatible endpoint. Full agent — 60 tools, 24 personas, 5-layer memory — with $0 ongoing cost. The floor is free, not a free trial.
BYOK keys live in your OS keychain. Never sent to our servers. Never logged.
See what she can do
Not another chatbot with a file reader. A full operating system for development.
Build with me
Describe what you want. 24 personas plan, verify, challenge, then execute. 60 tools. Full autonomy.
Learn anything. Free.
Curriculums, quizzes, spaced repetition, progress tracking. Five specialist personas teach you properly — not just dump information.
Audit everything
Five-persona security team. OWASP scanning, CVE research, exploitability verification. Type !! and let the auditors loose.
Free for everyone
1,500 credits every month — about a handful of coding sessions a day. No credit card. No account required. Teach Mode is discounted for everyone — education shouldn't be expensive. You only pay when you need more credits. Never for features.
Local first. Always.
Your data never leaves your machine unless you choose it. Memory, tasks, journal, settings — all local by default. Cloud sync is opt-in. No telemetry. No tracking. Your code is never used to train models.
Open source. Period.
Every line of code. The extension, the IDE, the companion, the CLI. Fork it, audit it, build on it. Trust is earned by transparency, not by marketing.
Six modes. One mind.
Type a prefix. Change how she thinks.
She forgets nothing
Five layers of intelligence. Not a chat log — a living understanding of who you are, how you think, and what you've decided.
Everywhere you work
We're not building another premium AI tool. We're building the one that makes the rest of them irrelevant.
Ready?
Install. Connect. She'll take it from there.
Open source. Local first. Community built. Always will be.