Build your own agent with VAF.

VAF as a Python library. One pip install, your choice of provider, tools as Python plugins. Written by you, loaded at startup.

Three lines to a running agent.

The Agent class is the stable surface of the framework. The same code runs a fully local GGUF model (nothing leaves the machine) or any supported API provider: veyllo, openai, anthropic, google, deepseek, openrouter.

✓

Streaming built in. Pass on_token and deltas arrive as they are generated.

✓

Conversations persist. save_session() writes to the standard session store; pass the id back to resume with memory intact.

✓

Your tools, discovered. Subclass BaseTool or ship a tool as its own pip package, and the agent picks it up at startup. MCP servers wire in the same way.

Read the embedding guide Runnable examples

quickstart.py

from vaf import Agent

# a local GGUF model, or any API provider
agent = Agent(config={"provider": "local"})

answer = agent.run("Summarize my inbox.")
print(answer)

# stream tokens as they arrive
agent.run("Explain async/await.",
    on_token=lambda s: print(s, end=""))

# persist and resume a conversation
sid = agent.save_session()
agent = Agent(config={"provider": "local"}, session=sid)

Install only what your agent needs.

The base install is the core runtime and provider SDKs. Heavy features are opt-in extras; tools whose extra is missing simply stay unloaded, and the agent still runs.

vaf[server]the HTTP and WebSocket API

vaf[desktop]the desktop window and tray

vaf[memory]long-term RAG memory (pgvector)

vaf[speech]offline speech-to-text

vaf[browser]browser automation tools

vaf[pdf]PDF extraction and OCR

vaf[docs]Office document tools

vaf[all]everything, parity with the full product

Your tools stay yours.

VAF is open source under the AGPL-3.0, with an explicit permission on top: plugins, tools and workflows that talk to VAF through its public interfaces may ship under a license of your choosing, including proprietary terms. For embedding VAF itself in a closed-source product or SaaS, a commercial license is available.

pip install, then build.

The embedding guide walks through tools, sessions, async use and packaging.

Read the embedding guide Get an API key