Agents Configuration

Everything under Settings → Agent and Settings → Memory. For the agents themselves — what they are and their capabilities — see the Agents feature page.

Defaults

• Default agent — Claude Code, Codex, OpenCode, or Cursor.
• Default model — the model used for new sessions of each agent. (The Claude model list and default ship with the app.) The list includes Claude Sonnet 5, which defaults to a 1M-token context window — on 1M-window models auto-compact stays suppressed (see below).
• Default reasoning effort — per agent (low / medium / high / extra-high, where the agent supports it).

Sandbox (Codex)

• Codex sandbox mode — danger-full-access (default), workspace-write, or read-only. This is sensitive: it changes what the agent can touch on disk.

Context auto-compact

• Auto-compact (default on) — the master switch. For Claude it enables the SDK's built-in auto-compact; for Codex and OpenCode the app sends /compact once usage crosses the threshold below. Suppressed on 1M-window models regardless.
• Auto-compact threshold (50–99, default 95) — the percent of the context window at which Codex/OpenCode sessions auto-compact. (Claude's threshold is fixed by the SDK and isn't affected by this.)

Sign-in & binaries

• Sign in with Claude — mints a long-lived claude.ai (Pro/Max) OAuth token so in-app Claude sessions use your subscription. Managed through the button in Settings → Agent (it's a secret, not a field you paste).
• Cursor API key — required for the Cursor agent; a Test connection button validates it.
• Agent binary paths — optional explicit paths to the Claude / Codex / OpenCode CLIs. Empty means auto-detect via PATH. Sensitive (it changes which executable runs).
• Extra parameters — per-agent CLI flags appended at launch. Sensitive (they can change auth, sandbox, or model behavior).

Third-party API providers

Claude Code can be pointed at Claude-compatible third-party endpoints. Each provider entry has an id, name, base URL, auth token, and optional model list and default/fast model ids. Built-in presets:

Provider	Base URL
Moonshot (Kimi)	https://api.moonshot.cn/anthropic
Zhipu (GLM)	https://open.bigmodel.cn/api/anthropic
DeepSeek	https://api.deepseek.com/anthropic

Set the active provider to one of your entries' ids to use it, or leave it empty to use the default Anthropic API. These settings hold auth tokens, so they're sensitive.

Memory

The memory system (see the Memory feature) is configured under Settings → Memory. Everything here is off by default.

• Distil session memories — master switch; pulls durable facts out of Claude/Codex transcripts. Gates the rest.
• Auto-distil on session end — distil automatically when a session closes.
• Dreaming (background consolidation) — merge duplicates, promote recurring patterns, prune stale entries; runs at most once per workspace+agent per day while no sessions are active. Can also be triggered manually.
• Memory agent — which backend (claude / codex) runs the reflection/dream passes, and the model for them.
• Codex memories — surface Codex's native memories in the Memories view and feed dreaming.
• Injection for OpenCode & Cursor — inject the workspace's app-owned memory into agents that have no native store.

Tasks

• Default task isolation mode — the git isolation seeded into new Task Board tasks: none (Current Branch), local-branch, or worktree. Configured under Settings → Task.