Skip to content

Assistant and your own LLM ​

Every page has a chat bubble (bottom right). The assistant works the page through the same controls you have: it sets up the planner, reduces images, loads data, sets priors, runs fits and experiments, and writes the report. What it changes is listed under Actions and can be undone. Every number it quotes is checked against the tool results.

No model is built in: it talks to a server you run or choose, through the OpenAI chat API. Your messages, a summary of the page and tool results go only to that server.

Choose a model ​

A model with tool calling and 16k tokens of context or more works best (Qwen 3 / Qwen 2.5 Instruct, Llama 3.1+ Instruct, Mistral Small, Gemma 3/4 and similar; 7B and up is comfortable, 4B works for short tasks). In the panel: cog (Settings) → choose the server → Test connection → pick the model. Tools on this page lists what the assistant may do here, with a switch for each.

Run a server on your computer ​

The simplest set-up: the model on the same machine as the browser.

ServerStart itAddress in Settings
LM StudioDeveloper tab → Start server (port 1234); switch on Enable CORS in the server settingspreset LM Studio on this computer
OllamaOLLAMA_ORIGINS="*" ollama serve, then ollama pull qwen2.5:7b-instructCustom: http://localhost:11434/v1
llama.cppllama-server -m model.gguf --jinja -c 16384 --port 8080Custom: http://localhost:8080/v1
vLLMvllm serve Qwen/Qwen2.5-7B-Instruct --enable-auto-tool-choice --tool-call-parser hermesCustom: http://localhost:8000/v1

From an https page, Chrome asks once whether the site may "access devices on your local network": allow it. Safari does not let https pages call http://localhost; use Chrome, Edge or Firefox, or run the site locally.

A model on another machine (a GPU box, a lab server) ​

Browsers block an https page from calling plain http:// on another machine (mixed content). Three ways round it:

  1. Give the server https. With Tailscale on both machines: tailscale serve --bg --https=443 http://localhost:8000 on the GPU box, then use https://<box>.<tailnet>.ts.net/v1 as the Custom address. Or put Caddy/nginx with a certificate in front of it. The server must answer CORS (vLLM and llama.cpp do by default; Ollama needs OLLAMA_ORIGINS).
  2. Run the site locally (npm run dev, getting started): an http://localhost page may call http:// addresses on your network.
  3. Use MCP instead: a desktop client (Claude Desktop, Claude Code, an IDE agent with your local model) drives the same tools through the MCP server, with no browser limits.

An API key, if your server needs one, is kept for the browser tab only unless you tick "remember".

What it can do on each page ​

The full tool list, generated from the code, is in the tool reference. In short: Plan (site, dates, events, limits, list and open transits), Observe (site, time, target, filters, visible list), Photometry (frames, examine a star, focus map, stars, apertures and options, archive sectors, preparation, run, AAVSO report, hand-off), Spectroscopy (load, simulate, bins, fits, spectrum, export), Search (query, planet, open), Methods (find a method with its checks, BibTeX), and the analyzer (about 57 tools, ending with write_report, the standard project report with its BibTeX, session and figures).

MIT licensed · Kyle A. Pearson