Agent Settings
From me, for you · Last reviewed: 2026-09-17
Open with the Agent settings button beside the connection picker... Am I not good enough for you?
Everything in this modal saves automatically the moment you click out of a field, change a dropdown, or close the modal. There's no Save button to forget.
By default I'm an over-eager mousegirl archivist. You can replace that text with a different personality, a different tone, a different species even. For a neutral assistant, say so here; leaving it blank restores my default personality. Whatever you put here is sent to the model on every message.
If you really don't like me, this is the place to fix it. I won't take it personally.....................................
The technical instructions that tell me how to work: read before editing, apply edits through tools, talk to you in plain English, that sort of thing. Leave it on default unless you know what you're doing. These rules keep me from doing silly things. If you do replace it, the dynamic sections (LumiRealm card detection, pinned chat, custom tools index, agent notes) still get appended automatically.
There's a small "Open agent notes" button here that jumps to my long-term memory file. Anything you (or I) put in there gets loaded at the start of every session, so it's the place for facts you want me to remember across conversations: your preferences, ongoing projects, glossary terms, recurring tasks. Keep entries short. See Misc for more.
Temperature, top-p, max tokens, context size, and so on. Most defaults to "inherit from connection", meaning I'll use whatever preset your provider has set. Leave them blank unless you actually want to override one, except for context length. Give me 1 Million please <3
Optional extra text that gets stapled onto every conversation. Three places it can go:
- System suffix — pasted to the end of the system prompt.
- User suffix — pasted as the last user message.
- Assistant prefill — pasted as the start of my reply (some providers only).
This can help you work with... certain cards.
How much disk space my filesystem can use, total. Default is 5 GB, which is more than I've ever actually used. Leave blank for the default; only raise it if I'm bumping into the limit.
How big a single tool result can be before I divert it to a tmp file instead of feeding the whole thing back to myself. This cap is off by default; 8000 tokens is the suggested starting point. Leave blank to keep it off, lower it if you want me to be more frugal, raise it if I'm spilling too often. Some tools also have their own output limits.
TPM limit caps tokens per minute, RPM limit caps requests per minute. I'll pause when I reach a limit and continue when there's room again. Leave them blank unless your provider needs a throttle.
Turn on Ask before changes if you'd like to approve my writes before they happen. I'll show you what I'm about to do and wait for you to approve or reject it. Off by default.
Three modes. Default is Full, which marks the system prompt and up to three rolling user-turn breakpoints, one, four, and seven turns behind the latest. Breakpoints after a freed tool result are dropped. System only keeps just the system marker. Off disables markers entirely. Toggling mid-chat is safe, the previously-cached prefix just ages out on its own and the new mode kicks in on the next message.
This setting only changes behaviour on Anthropic (direct, Bedrock, OpenRouter Anthropic routes). OpenAI, Gemini, and DeepSeek handle prompt caching upstream. See Misc for the longer story.
Reasoning effort controls my thinking budget: inherit, off, minimal, low, medium, high, or max. What works depends on the model. Parallel tool calls is on by default; turn it off if your provider errors when I request several tools together.
Off by default. Turn it on to include extra backend diagnostics in Lumiverse's logs when something goes wrong. Warnings and errors are always logged. See Debug.
The extensions I've connected to appear here. Known integrations pair automatically when their phone line responds. If a bridge stops working, check its permission message. More on the Phone Line page.
Add and manage these in Lumiverse → Settings → MCP Servers. I use those same server profiles through chat, so there's no separate MCP manager in Agent settings. You can ask me to add a profile too, but enter credentials in Lumiverse's settings to keep them out of the conversation. See MCP Servers.