Settings

Last updated 2026-08-21

⌘, opens it. Eight tabs, mixing things you set once with things you may never open.

TabWhen to open it
AppearanceOnce — theme, text size, menu bar
LLMOnly for AI features — keys, models, fallbacks
ocul-pmManaging agent rules, plugin install instructions
Indexing & RAGWhen search isn't behaving as expected
GraphWhen the code map is too dense or too sparse
DataExport and cleanup
DiagnosticsWhen reporting a problem
UpdatesManual check, past release notes

Appearance

Theme

A two-layer system: Light / Dark / System, with preset themes on top.

  • 5 presets — Solarized · Sepia · Nord · Dracula · High Contrast
  • 6 accent colors — Green (default) · Blue · Purple · Orange · Rose · Teal
Note

Choosing a preset disables accent selection. Presets bring a whole color system of their own. To change only the accent, stay on Light/Dark/System without a preset.

Two text sizes

App scale and terminal font size are separate.

UnitApplies to
Text sizeScale (%)The whole app — like browser zoom
Terminal font sizepxTerminal screen, dock, detached windows

The terminal is in px because a fixed-width grid needs px precision more than a percentage. Inside the terminal, ⌘+ / ⌘− adjust it and ⇧⌘0 resets.

Two languages

UI language (on-screen text) and AI writing language (the language of entries, planner items, and retros the AI produces) are separate. You can run an English UI and keep records in Korean. Already-written documents don't change.

Menu bar residency and external editor wiring are covered in Windows, Tabs, Terminal, Code.

LLM

Needed only for AI features (retro narrative generation, Agent panel chat). Journaling itself needs no key.

API keys

Stored only in the OS keychain — never in the app database or a file. The status (set / unset) is cached locally, so opening Settings doesn't trigger a keychain prompt; press Verify against keychain only when you really want to check.

Providers and models

Pick a default provider, and optionally override the model per provider. Leave it blank to use the built-in default.

  • OpenRouter — hundreds of models over an OpenAI-compatible API. Model id example: openai/gpt-4o
  • NIM — the OpenAI-compatible endpoint at integrate.api.nvidia.com

Fallback chain

When the default model call fails, these are retried top to bottom. One provider:model per line.

openai:gpt-4o-mini
anthropic:claude-3.5-haiku-latest
openrouter:openai/gpt-4o
Tip

If you run retros during hours when rate limits bite, a fallback chain makes a real difference.

Generation

Temperature (lower is focused, higher is creative), max output tokens, and a system prompt prepended to every chat.

Indexing & RAG

The tab to open when code search isn't behaving.

  • Auto-index on change — when files change, only the changed files are indexed. To rebuild from scratch, use Rebuild index
  • Chunkinglarger chunks mean richer context per snippet, smaller ones mean sharper search. Overlap has to be smaller than the chunk size
  • RAG context — how many top chunks go to the model per chat message
  • Work context — automatically inject recent entries and the AGENTS rules into chat so direction survives a change of session or model. More entries means richer context but more tokens (0 injects rules only)
  • File scan — max file size, extra exclude patterns (gitignore-style, applied on top of .gitignore)
Warning

Scan setting changes take effect at the next reindex. Rebuild the index to apply them now.

Graph

Defaults for the code map — whether isolated files show by default, and the auto-group threshold (directories with at least this many files become collapsible groups).

In a big repo, lower the threshold when the map is too dense; raise it when everything is lumped together.

Notion export

Exports retros and artifacts as new pages under a parent page you choose. Connect through the browser with your account, or paste an internal token. That token also lives only in the OS keychain.

Diagnostics · Updates

  • Diagnostics — the information to attach when reporting a problem. DevTools opens from here too
  • Updates — manual check and past release notes (GitHub releases). Normally the app updates itself
Note

An update restart relaunches the Claude Code adapter along with the app. That's why restarts are deferred while an answer is streaming — the conversation is on disk and comes back, but the answer in flight at that moment is lost.

ocul-pm

The tab for managing agent rules — choose which agents' rules files to install, use Resend rules when the spec has moved, and Detect to narrow to agents that appear to be in use here. The plugin install commands and the Claude Desktop MCP snippet live here too, with copy buttons. (Details)

Next steps