Settings reference

Settings has fifteen sections, listed down the left side. Each one below gives the main options, their defaults and the page with more detail.

Open Settings

Click the gear button at the right of the header; click it again, or ← Back, to close it. There is no shortcut for it. Settings remembers the last section you opened.

General

Appearance and how replies are formatted.

• Theme: System (default), Light or Dark.
• Highlight color: the accent color. Default Teal.
• Response Format: Minimal, Standard (default) or Rich.

See the chat page.

Inference

Where the model runs, which model is used and how big the context is.

• Inference backend: Local (Haruspex-managed) (default), Remote server (advanced) or OpenRouter (cloud).
• Models: download and switch local models.
• Context Size: 32K by default (8K to 256K).
• Let models use system RAM: off.
• Parallel streams: 1 (also 2 or 4). More lets a background job run beside chat. Each stream gets the full context size, so it needs more memory (larger sizes may grey out), and replies slow while two run.
• Extra llama-server arguments: empty. Added after Haruspex's own arguments, so they override them. Takes effect when the server restarts.
• Server: status, port and Restart Server; Run Setup Wizard.

See the models page.

Agent

How the model reasons and how long its replies may be, the Python sandbox, and pictures in answers.

• Reasoning mode: on. Reasoning effort: medium.
• Additional system prompt: empty.
• Max response tokens: 8192. Max response tokens (file writes): 65536.
• Enable Python sandbox: on. Approval prompt: Once per chat. Network access: Internet and local network. Execution timeout (seconds): 60.
• Pictures in answers → Include images in answers: on.

See the chat page.

Memory

What the assistant remembers between chats.

• Carry facts and preferences from one conversation into the next: on. It needs a small embedding model (about 65 MB), downloaded only when you ask.
• Ask before saving something you asked me to remember: on.
• Find duplicates: merges memories that say the same thing, after you approve each.
• A list of everything remembered, to edit or delete.

See the memory page.

Skills

Instructions the assistant can load for specific tasks.

• When the model uses skills: Automatic (remote models only).
• Skills: turn single skills on or off; Open skills folder.
• Other skill folders: extra folders to read skills from.
• Repos: whether the Shell assistant and Code sessions use a repo's own skills and AGENTS.md.

See the skills page.

Audio

Voice input and spoken replies.

• Text-to-speech voice: Heart (Female).
• Read tables by subject: on.
• Audio output (TTS playback) and Audio input (microphone): pick a device; Refresh rescans.

See the chat page.

Search

Which provider answers web searches.

• Search provider: Auto (rotates free engines). Also DuckDuckGo, Brave Search (API key required), SearXNG (self-hosted) and Browser-assisted.
• SearXNG Instance URL: http://localhost:8080.
• Result recency: Any time.

See the search-and-network page.

Network

Proxy settings for network connections. Proxies never apply to this computer's own services.

• Network Proxy: None (default) or Manual.
• Web Search Proxy: Network proxy (default), None or Manual.

See the search-and-network page.

Integrations

Email, calendar and contacts, and MCP servers.

• Email: Add email account. Any IMAP account; most need an app password.
• Calendar & Contacts: Add a calendar link or Add a server account (CalDAV/CardDAV).
• MCP integrations: Add from the catalog, Add a server on this computer or Add a server on your network.

See the integrations page.

Screen

Whether the assistant can see your screen.

• Let the assistant take a screenshot when you ask it to: off.

See the integrations page.

Shell

The Shell tab's terminal and its assistant.

• Shell binary: blank, which uses your $SHELL.
• Recent shell commands attached to each chat message: 3.
• Include your shell history file in prompts: on.
• Max output bytes per captured command: 8192.
• Start new shells with Full access: off, so new tabs start Read-only.
• Command execution: Auto (the terminal when capture works, else a separate process), for Full access.
• Memory limit: 50 % of RAM.

See the shell page.

Code

How the coding agent runs commands, in the Code tab and in the Shell assistant with Full access.

• run_command timeout: 30 seconds.
• Max steps per task: 40.
• Background log size: 5 MB per process.
• Auto-approve commands: off.

See the code page.

Image

Where pictures are generated, if anywhere.

• Backend: None (default), ComfyUI or Bundled engine. Nothing is downloaded or started until you pick one. Bundled engine is hidden where it can't run.

See the images page.

Guest chat

Let people on your network chat with this Haruspex.

• Let people on your network chat with this Haruspex: off. Traffic is not encrypted.
• Port: 8787.
• Once it is on: a link and QR code to share, and a list of guests you can disconnect.

See the guest-chat page.

Remote control

Use your Code sessions from your other computers. Off; port 8788.

See the remote-control page.

Feedback

Report a bug or ask for a feature.

• Open feedback issue…: opens a filled-in GitHub issue in your browser. API keys are removed, but check it before you send it.
• Save Full Diagnostics…: saves logs to a file you can attach to the issue.

See the troubleshooting page.