01
For power usersDesktop app
Install it, sign in, and the agent is in its own window.
- Your data and keys stay on your machine
- Local models, or one API key for everything
- Schedules, memory and tools built in
The local-first framework for agents that learn.
The open-source AI agent with persistent memory and tools for web, code and files. Fully self-sufficient in-house, or with models from the cloud.
18,700+ downloads·Open source, AGPL-3.0·macOS, Windows & Linux
Three ways to run
Local-first by design. You decide where VAF runs.
VAF runs on your machine, on your own server, or as a framework inside your code. Memory and keys live where the agent runs.
02
For teams & companiesSelf-host it
VAF runs as a host on your own server, reached at your own address instead of through a hosted platform.
- One host serves the whole team over HTTPS
- Separate accounts with isolated memory
- No per-seat subscription
03
For developersBuild on the framework
One import brings the whole agent, with memory, tools and sub-agents. Your own tools are plain Python files; MCP servers plug in on top.
- One pip install, provider-agnostic
- Your own tools via BaseTool or MCP
- Ship plugins under your own license
See it work
The main agent
You ask in plain language, and the main agent plans the work and delegates each part to the specialist sub-agents below.
Why Veyllo
Hand work over instead of operating a tool.
The agent learns your tools, reads its way into your documents and keeps going while you do something else. On your own machine.
Reads itself in, and keeps it
Whole documents go into long-term memory, contracts and reports of several hundred pages included. What the agent learned from them is there in every later conversation, not only the one the file arrived in.
Works around the clock
At set times or continuously in the background, unattended. Once at noon on the 23rd, every day at 5 pm, or without stopping. The result is ready when the run ends.
Learns your tools
The agent teaches itself new tools, custom ones included: it runs them, reads the results and adjusts until it can use them on its own. Every run makes it surer of them.
Your machine, your server, your call
Local model, your own server, or one API key for every provider. You decide where the data lives and which model does the work. No per-seat subscription.
Splits large work up
The agent breaks a job into steps and hands off whatever can run in parallel. An agent-to-agent protocol lets it work alongside further instances, and alongside agents from other providers such as Claude, Codex or OpenCode.
Local-first AI agents
What it means to run an agent on your own machine, and why local-first matters for privacy, cost and control.
Read the explainer →Build on VAF
Your own agent, three lines in.
The desktop app runs on this framework, and you build on the same primitives. Install it from PyPI and ship your own agents, tools and products on top.
✓
Provider-agnostic. The same code runs a local GGUF model or any API provider; swap with one config line.
✓
Bring your own tools. Subclass BaseTool or wire in MCP servers, and the agent picks them up at startup.
✓
Your license, your product. Plugins, tools and workflows built on the public interfaces may ship under your own terms.
Your machine. Your data.
Your agent.
Veyllo API · preview
Every model VAF needs, behind one key.
VAF needs models for text, vision and speech. Normally that means an account with every provider, every key wired in by hand, a separate invoice per provider and no combined view. The Veyllo API bundles that into one key and one bill.
✓
One key instead of many. Every model VAF uses, behind a single API. No separate accounts or keys at each model provider.
✓
No models to host yourself. No GPUs, no self-run model servers. Veyllo handles compute and model hosting.
✓
Built for production. Keys, rate limits and usage metering from day one.
Endpointapi.veyllo.app/v1
Compatible withOpenAI SDKs
Modelschat · vision · transcribe
Pricingpay as you go
Chat
€0.90 in · €1.90 out
per 1M tokens
Vision
€3.45 in · €20.70 out
per 1M tokens
Transcribe
€0.012 / min
billed per second
Install
Bring the agent to your desktop.
Two steps: clone the repository and run the installer. You pick the platform; VAF sets up the rest: container runtime, Python and Node.
Open development
Follow the build on GitHub.
Read the code, follow the commits, and raise issues that reach the developers directly.
30
stars on GitHub
18,700+
downloads and counting
Active
development
1 dot ≈ 2 agents deployed
Growth since the alpha · Jun 26, 2026
Stars30
Downloads18,700
Data was not previously recordedData Fetching
Jun 26Aug 6Sep 15
News
Published recently.
VAF 0.1.0a26: lower cost per request through prompt caching
VAF's daily spend limit stopped too late because the cached share went uncounted on Anthropic; 0.1.0a26 counts it.
VAF 0.1.0a24 and 0.1.0a25: hands-on control of the browser
A fresh install of 0.1.0a24 never gets past the splash screen; 0.1.0a25 builds the interface and frees the next update too.
VAF 0.1.0a23: joining a room without a VAF install
In 0.1.0a23 a foreign client joins with one standard-library file; the sha256 and CA fingerprint in the invitation are what verify it.
Ready for your agent?
The desktop app installs everything it needs; the Veyllo API supplies the models.