xng-agent

A NOOA agent that researches a topic using live web search (xng / SearXNG) and real page browsing (browser-use primitives). No memory, no browser sub-agent: the only LLM loop is NOOA's.

The agent is a single Python class (src/xng_agent/__init__.py); the model-facing surface:

• await search_web(query, limit) — deterministic; wraps the xng Python API in a worker thread (asyncio.to_thread) so the blocking HTTP call can't stall the browser loop.
• browse(url) — deterministic; opens the URL in a shared headless Chrome and returns the page's visible text (truncated to 3000 chars) and title. Drives browser-use session primitives directly (CDP navigate + document.body.innerText), so no nested LLM loop.
• research(topic) — agentic (CodeAct, ... body): the model writes Python, searches, browses, and returns a validated ResearchReport (summary, cited key facts, sources).
• today — state field (weekday, date, local time + UTC offset) visible to the model so it can resolve relative dates like "this weekend" to concrete dates.

Prerequisites

• A llama-server router (OpenAI-compatible). Defaults: http://127.0.0.1:8080/v1 with model Qwen3.6-35B-A3B-MXFP4_MOE; override with NOOA_MODEL / NOOA_LLM_BASE.
• A running SearXNG instance (configured for xng, e.g. SEARXNG_URL).
• A local Chrome at /usr/bin/google-chrome (override with CHROME_PATH).

Usage

Starting the nooa dev console first and leaving it open makes it easy to review the agent traces.

export NOOA_VIEWER_AUTH_TOKEN=$(openssl rand -hex 16)
uv run nooa start-dev -h 0.0.0.0

Then run the agent

uv run xng-agent "subject to search"

Notes

• One shared Chrome session per run: started lazily, and browse() waits until the CDP connection is actually open before touching a page (browser-use's start() returns before the launch handler finishes). Every page operation is timeout-bounded; a hung or crashed session is killed and replaced on the next call.
• The session is torn down with kill() at the end of the run. browser-use's stop() deliberately keeps the browser process alive (for reconnecting) — using it leaks Chrome.
• browse() returns the final URL after redirects and reports empty content honestly for pages that only render via JS; the model is expected to skip those.
• Thinking is disabled for the NOOA loop via chat_template_kwargs (litellm extra_body) for speed.