Run Ava for $0 on your own machine
The whole agent — every mode, every tool, your memory — running on a local open-weight model via Ollama or LM Studio, for zero ongoing cost. No key, no metering, nothing leaves your device. Here is exactly how.
Most AI tools have a "free tier" that is really a trial — enough to get you hooked, then a paywall. Ava is built the other way round: the floor is genuinely free, forever, because you can run the entire agent on your own hardware and pay us nothing at all. This is the cheat code. Here is exactly how to use it.
What you actually get
Not a stripped-down "lite" mode. The full agent:
- Every mode — Code, Plan, Chat, Teach, Security, Brainstorm, Write — and every tool.
- Your local-first memory, so Ava still learns and remembers you across sessions.
- Zero ongoing cost. No credits, no API bills — you already own the hardware.
- 100% private: the model runs on your machine; your prompts and code never leave it.
- No rate limit but your own silicon.
What you need
- A machine that can run a local LLM. A modern laptop handles small-to-mid models; more RAM or VRAM lets you run larger, more capable ones.
- A local model runner: Ollama (simplest), LM Studio (nice UI), or vLLM for power users.
Step 1 — Get a model running locally
With Ollama, install it and pull a capable, tool-call-friendly model. Coding-tuned models like Qwen2.5-Coder are a good starting point; larger variants are noticeably more reliable for agentic work.
ollama pull qwen2.5-coder
ollama serve # usually already running in the backgroundPrefer a UI? LM Studio does the same job — download a model in the app, then start its local server from the Developer tab.
Step 2 — Point Ava at it
In the extension or the desktop IDE, open Settings → Custom Model and enter two things. No API key required:
- Base URL — Ollama: http://localhost:11434/v1 · LM Studio: http://localhost:1234/v1
- Model name — exactly as your runner lists it (e.g. qwen2.5-coder).
Step 3 — Use it
Select your custom model and start working. Same chat, same modes, same memory — now running entirely on your hardware, at $0. Ask it to explain a file, build a feature, or teach you something: the agent loop is identical to the managed models.
An honest word on model size
Local open-weight models are weaker than frontier cloud models at reliable tool-calling and long, multi-step tasks — the smaller the model, the more that shows. Ava’s orchestration (role-specialised sub-agents, verification and repair loops) is deliberately designed to get real capability out of modest models, but set expectations honestly: a 7B model is great for chat and small edits; a larger model, if your hardware allows, is far better for autonomous, multi-file work. If a task struggles locally, that is the model, not the agent.
Why we tell you this
Every other platform hides the free path. We lead with it, because the point of Ava is access — capable, private, agentic AI that anyone can run regardless of budget. The managed plans exist for people who want the convenience and the frontier models, and those subscriptions are what keep the free tier and the open source alive. But the floor is genuinely free, not a free trial. That is the whole idea.