Local private AI assistant that browses the web, writes code, and plans multi-step tasks

What they did

They cloned the AgenticSeek repository and stood up a fully local, voice-enabled agent stack on their own machine. Instead of pointing at a cloud API, they wired the assistant to a local reasoning model served through llama.cpp, keeping every file, search query, and conversation on-device. They configured the Python 3.10 environment, installed dependencies, and launched the web UI. Then they exercised the agent's core loop: asking it to autonomously browse the web (search, read pages, extract information), write and execute code in Python and other languages, and decompose a larger request into a multi-step plan handled by specialized sub-agents. They also tried the smart agent selection, letting the system route each subtask to the right agent automatically, and experimented with the speech-to-text voice interface. The whole exercise was about validating a private, zero-cloud alternative to hosted agent products like Manus, and seeing how far a local model could carry a real multi-tool workflow without leaking data off the device.

How it went

The stack came up and handled browsing, coding, and planning tasks end to end without any cloud calls, which is the headline win for privacy-sensitive work. Quality tracks the local model: smaller models stumble on long plans and tool routing, while larger ones behave more reliably. Voice is functional but still marked in progress, and the project itself is a side-project with no roadmap or funding, so expect rough edges and manual troubleshooting rather than a polished product.

Worth knowing

You need real hardware — a capable GPU and enough VRAM for the chosen local model — plus Python 3.10.x and Git. Model choice heavily affects reliability; weak models produce broken plans and failed tool calls. Voice features are incomplete. There is no official roadmap or support, so setup issues are largely self-serve. Web browsing and code execution give the agent real reach, so sandbox or review its actions before pointing it at sensitive files or credentials.

Wiring

Order matters: install Git and Python 3.10.x first, clone the repo, then create and activate a virtual environment and install the Python dependencies. Next, serve a local reasoning model through llama.cpp (build it, download a GGUF model, start the server on a local port) and set AgenticSeek's configuration to point at that local endpoint instead of any cloud provider. Launch the AgenticSeek app/UI, then verify the agent loop: the planner decomposes a request, the router picks a browsing, coding, or other specialized agent, and tools execute locally. Keep the model server running in one process and the agent in another, and confirm no outbound API keys are configured so traffic stays on-device.

Try it yourself

Set up AgenticSeek in this repo as a fully local agent. Steps: (1) verify Python 3.10.x and Git are installed; (2) create a venv and install requirements; (3) start a local model server with llama.cpp and point AgenticSeek's config at it; (4) launch the web UI and confirm it loads; (5) run three smoke tests — a web-browsing research query, a small Python coding task it writes and executes, and a two-step plan that requires agent routing. Report which tests pass, note any config values I must change for my hardware, and flag anything that would send data off-device.