Skip to content

LLM Assistant (Experimental)

An optional chat panel for questions about the loaded project and NeuroSeq. It receives the documentation, a fresh project summary with each message, and read-only access to project details and the NeuroSeq source. Behaviour and supported providers may change.

Setup

Open Settings → LLM Assistant, choose a provider and model, then click Test connection (checks sign-in without sending a prompt).

Provider Requirements
Codex (local sign-in) Codex CLI and codex login. Recommended.
Claude Code (local sign-in) Claude Code and claude auth login (subscription, Console account or ANTHROPIC_API_KEY). Recommended.
Ollama (local) (default) running Ollama daemon (default http://localhost:11434) and a pulled model (default qwen2.5:7b)
OpenAI API key; any OpenAI-compatible endpoint via the base URL field
Anthropic API key
  • Ollama, OpenAI and Anthropic need the chat extra: pip install "neuroseq[chat]" (installs openai, anthropic, httpx, keyring).
  • API keys are stored in the OS keyring. Without a keyring backend, NeuroSeq reads OPENAI_API_KEY / ANTHROPIC_API_KEY.
  • Codex and Claude Code: sign in under the same OS account as NeuroSeq. Leave the executable field empty to use PATH (restart NeuroSeq after changing PATH); Codex may also point to the binary bundled with its VS Code extension. Model default uses your CLI configuration; Claude Code also offers opus, sonnet, haiku, Codex lists models via Refresh.

Use

Toggle View → Show Chat (robot icon). Enter sends, Shift+Enter adds a line, Stop cancels. Replies stream in.

The assistant can read model equations, parameters, initial conditions, synapses, monitors, runtime state, input arrays, experiments, imaging metadata (large arrays in pages) and Python files inside the installed neuroseq package. It cannot read other files, run code or run simulations. Anything it reads is sent to the provider.

Data scope

The Don't share project data with the provider checkbox in Settings → LLM Assistant limits what the assistant can read:

Checkbox Readable
On Docs, source, model, parameters, runtime state, neuron count and duration
Off All of the above, plus input arrays, experiments, Cellpose and imaging data and their names

Each provider keeps its own value. Default: on for OpenAI, Anthropic, Codex and Claude Code; off for Ollama, which runs locally. Turn it on for an Ollama server on another machine. The chat badge shows "data not shared" while it is on.

The checkbox only limits what NeuroSeq hands to the assistant. Codex keeps its own shell, file and MCP tools (it has no single switch that turns them all off), and its read-only sandbox can still read any file your user account can, including saved projects and recordings. The settings dialog warns about this when Codex is selected, and the chat badge then reads "NeuroSeq data not shared". To keep project data private, use OpenAI with an API key (or any OpenAI-compatible server, such as Azure OpenAI or a local vLLM, set as Endpoint), Claude Code (all native tools disabled) or Ollama.

Earlier replies can contain project data, so the chat clears its history (with a short note) when you switch provider or endpoint, or turn the checkbox on. Changing only the model keeps the history. Every chat panel, including detached ones and other project tabs, switches to the new settings as soon as you change them.

Writing analysis plugins

Ask e.g. "Create an analysis plugin that plots the mean trace across selected neurons." NeuroSeq shows the target path and full code (or a diff) and saves only after you click Save plugin. Review the code: a shared project could contain text that steers the assistant towards harmful code.

  • Saved only to the per-user plugin folder (Linux: ~/.local/share/neuroseq/analysis_plugins, override with NEUROSEQ_USER_PLUGIN_DIR), never into the package.
  • Names starting with _ or matching built-in plugins are rejected; only plugins already in your folder can be overwritten.
  • Syntax is checked, correctness is not. Reopen the right-click analysis menu to load the plugin (this imports its code).

Limitations

  • Up to 12 read/write requests per turn; ask a follow-up to continue. Reliability depends on the model following NeuroSeq's JSON request protocol.
  • Conversations are not saved between sessions.
  • Codex runs each request in an ephemeral read-only sandbox, but your configured Codex tools still apply and its filesystem sandbox is not limited to the NeuroSeq package. Claude Code runs claude --print in an empty temporary directory with all native tools, MCP servers and settings disabled.