Cogtrix quickstart

Three commands and you’re talking to a local LLM with 67 tools at its disposal.

The three commands

ollama pull qwen3:8b            # any GGUF model works
uv sync
uv run python cogtrix.py

That’s it. Cogtrix auto-detects Ollama on localhost:11434, loads its default config, and drops you into an interactive prompt.

Your first prompt

You> Find the five most-cited deep-learning papers from arXiv in 2025,
     summarize each in two sentences, and save the list to papers.md.

What happens next:

1. The agent classifies the request as a research task.
2. It calls http_get against Semantic Scholar for citation counts.
3. For each paper, it pulls the abstract from arXiv.
4. It composes the summary and asks for confirmation before writing papers.md to disk.

You answer y to confirm the file write, and the file appears in your cwd. One prompt, five tool calls.

What to try next

• /help — list every slash command.
• /think <task> — engage the Tree-of-Thought engine for tasks that need real reasoning.
• /tools — see what’s available; load extra tools with /tools load <name>.
• /model gpt-4.1 — switch to a cloud model (requires the right *_API_KEY env var).

→ First session walkthrough — a longer example showing memory + tool confirmations in action.