Settings Reference
Every tab in the Cerevisor Settings popup, every field, every toggle, what it does and what the default is.
Open the Settings popup from the gear icon in the title bar. It has 17 tabs, listed below in the order they appear in the tab bar.
This page is the master reference for every tab: its purpose, the main fields it contains, and notes on what each setting actually changes when you flip it.
| Tab | What it covers |
|---|---|
| Workflow | Workflow name/description, prompt mode, preamble, thinking mode, run options, spend caps, wow + situational-awareness directives, custom suffix. |
| Permissions | Approval mode, the team-wide tool-category caps, web access scope, Outlook backend. |
| Input | What input the workflow accepts when run — none, user message, file, or folder. |
| Agents | Role palette ordering and visibility, custom agent templates. |
| Chat Builder | The chat builder's model override and intent-extraction pre-pass, global and per-workflow. |
| Integrations | Model providers, smart routing, context & compaction, per-workflow provider override, Cursor cloud execution. |
| Mobile | The mobile companion: enable, pair a phone, manage paired phones. |
| Skills | Skill folders, semantic skill retrieval, re-embedding. |
| Automation | Skill auto-authoring, background model & failover, prompt caching, run-history retention, and the Operator's reach. |
| Hooks | User-configurable lifecycle hooks that watch or block runs. |
| Learning | Welcome intro, sample folders, contextual guidance cards. Toggle the whole layer on/off. |
| MCP | Add/remove/test MCP servers. |
| Web Search | Default web-search provider (Tavily, Perplexity, Linkup), per-provider API keys, test connection. |
| Vault | Secrets your workflows can use (API keys, tokens) without exposing them in .cerevisor files. |
| Subscription | License state, trial countdown, manage subscription, upgrade. |
| Updates | Current version, check for updates, last-checked time. |
| Reset | Selective factory reset of local data. |
Workflow
The home tab. Settings that affect the active workflow's identity and execution behavior.
| Field | What it does |
|---|---|
| Workflow Name | The workflow's display name. Editable here or in the title bar. |
| Description | A short description of what the workflow accomplishes. |
| Prompt Mode | The "Structured orchestration prompt" toggle. When on, the generated prompt is enriched with subagent types, model hints, tool scoping, and structured handoffs. Independent of the runtime you actually execute with. |
| Custom Preamble | Free-form text that overrides the auto-generated preamble of the prompt. |
| Default Thinking Mode | Applies to all agents whose thinking mode is set to Inherit. Quietly ignored on models that don't support thinking. |
| Options | A group of run-behavior toggles: include permissions request, include deliverables section, auto-flow files between agents, strict output contracts, halt on contract failure, require user input before running, confirm before resuming from an agent, runtime skill loading, memory mode (read + write / read only / off), subagent turn limits, approval + inactivity timeouts, rate-limit auto-wait, spend caps (per workflow / per agent / per agent-chat), and cost optimizations. |
| Custom Suffix | Text appended to the generated prompt after everything else. |
| Wow Directive (global) | Whether Cerevisor's signature quality directive is applied to every prompt and every agent. One global toggle. |
| Situational Awareness (global) | Whether the situational-awareness directive is injected into every workflow. Pushes agents to surface uncertainty and name unknowns explicitly. |
| Super Controller Mode (global) | An additional global directive layer for stronger orchestrator control. |
| Output Versioning (global, experimental) | Version output files on every run instead of overwriting them. |
| Files Cerevisor writes to your project | A read-only list of the only files Cerevisor itself will ever create in your project folder — everything else lives under ~/.cerevisor/. |
| Command Bar | Expanded view toggle: larger agent tabs with labels instead of compact icons. |
| Keyboard Shortcuts | A read-only cheat sheet of the main shortcuts. |
Permissions
The team-wide caps enforced by the runtime while workflows run. They apply in every approval mode, including Hands-free. Per-agent tool access (set inside each agent) can only narrow these limits, never broaden them. Agents that run on an external tool (Claude Code, Codex, Cursor, Antigravity) receive these limits once at launch.
| Field | What it does |
|---|---|
| Approval mode | How often Cerevisor pauses for your sign-off: Inherit (use the app-wide mode), Oversight (approve each action), Auto (runs by itself, asks before consequential actions), Hands-free (never asks; requires a one-time confirmation the first time you pick it). |
| Tool categories | Individual on/off caps: file creation, file modification, bash execution, web access, git operations, package installation, and external services (Outlook mailbox/calendar — drafts only, agents can never send mail). |
| Web access scope | Visible when web access is on. How far web_fetch may reach: Approved sites only (recommended), Anywhere, or Off (web_search stays available). Part of the always-on protection against malicious web content. |
| Outlook backend | Microsoft Graph (cross-platform, OAuth-permissioned), Local Outlook (drives classic Outlook desktop), or Auto. |
| Prompt export | Whether exported prompts (CLAUDE.md / Markdown) restate the caps above as instructions. Does not affect runtime enforcement — caps are always enforced. |
Input
Defines what feeds the workflow when it runs.
| Input type | What it does |
|---|---|
| None | The workflow runs with no input — agents start from their instructions only. |
| User Message | A text prompt prepended to a designated entry agent's prompt at run time. |
| File | A single file attached to the entry agent's inputs. |
| Folder | Opens the folder in a left-side file explorer panel; drag any file onto an agent's INPUT zone to attach it as context. |
| Webhook | Shown in the picker but marked Coming soon — not selectable yet. |
For User Message and File you also pick the Entry agent: the agent that receives the input.
Agents
Manage the agent palette and your saved custom agent templates in one drag-reorderable list.
- Reorder roles and templates by dragging; the order sets the palette and the Shift+1–9 / Shift+0 placement shortcuts.
- Hide built-in roles you don't use — hidden roles disappear from the palette. A reset control brings the defaults back.
- Custom agent templates appear in the same list; you can edit or delete them, and drop one into the active workflow.
Chat Builder
Configure the chat builder you open with the chat panel.
- Model — which provider/model the builder chats with. Only providers with a chat surface (Anthropic, Gemini, OpenAI-compatible) can drive the builder directly; if your app default is Claude Code, Codex, Cursor, or Antigravity, the builder can still work through it one task per message when that option is enabled.
- Intent extraction — an optional pre-pass before each new builder session: one extra model call that summarizes your request and ranks the available skills and
.mdfiles by relevance, so the builder makes better matches. - This workflow overrides — override the global chat-builder settings for the current workflow only.
Integrations
Where your model providers live.
- Providers — the provider library: add, edit, enable/disable credentials, and pick the app-wide default. See Managing the provider library.
- Smart Routing — auto-route agents to the best model per task, with a strategy of Cost-optimized, Balanced, or Quality-optimized.
- Context & Compaction — how the orchestrator manages long conversations and context limits.
- Default provider for this workflow — an optional per-workflow override so one
.cerevisorfile can pin a different provider without changing the global default. - Cloud execution (Cursor only) — run agents on Cursor's cloud instead of your machine; shows guidance when a different provider is configured.
Mobile
The desktop half of the mobile companion.
- Mobile companion — the master toggle: "Allow your phone on this LAN". Off by default; opens an encrypted local listener only while Cerevisor is running. You can pick which address phones connect to (e.g. your own VPN — Tailscale, WireGuard — for access away from home).
- Pair this desktop — shows a two-minute, single-use QR code to scan with the Cerevisor Companion app. The desktop confirms the phone before saving it.
- Paired phones — the list of paired devices with last-seen times and a Revoke action.
- Connection facts — plain-language notes: there is no Cerevisor internet relay, traffic never touches Cerevisor's servers, and the desktop app must remain open.
See Mobile companion security for the transport and trust model.
Skills
| Field | What it does |
|---|---|
| Skill folders | Cerevisor always scans ~/.claude/skills/ and ~/.cerevisor/skills/; auto-discovered CLI tools appear read-only. You can add your own extra folders — changes trigger a live re-scan, no restart needed. |
| Semantic skill retrieval | Whether Cerevisor uses on-device embeddings to rank skill relevance (nothing leaves your machine). |
| Re-embed all skills | Button. Triggers a full re-embed pass over every skill. Run after large skill edits. |
Automation
Background behaviors that run around your workflows.
Skill auto-authoring
When Cerevisor sees the same error, success sequence, or user correction repeat across several runs, it can draft a skill capturing the pattern. Drafts land in your Skill Workshop for review before becoming visible to future runs.
| Field | Default | What it does |
|---|---|---|
| Enable skill auto-authoring | On | Runs once per workflow after its retrospective pass. |
| Occurrence threshold | 3 | How many runs a pattern must repeat before a draft is written (2–10). |
| Cooldown (sessions) | 5 | How many runs to wait before re-drafting the same pattern (1–30). |
| Auto-activate drafts | Off | Skip review — drafts go live immediately. Power-user option. |
Background model & failover
Informational: background features (memory insights, retrospectives, skill summaries) run on your default provider — on a subscription default (Codex, Cursor, Claude Code, Antigravity) they cost nothing beyond the plan you already pay for. If the default fails at runtime, background work falls back to another chat-capable credential in your library; that fallback spend is reported under Analytics → Models.
Prompt caching
Anthropic-only: cache lifetime (5 minutes or 1 hour) and pre-warm. Other providers ignore these.
Run history retention
How many past runs to keep on disk: 200 (default), 1,000, 5,000, or 20,000.
Operator
How far the Operator — the assistant that works on your workspace in the background while the app is open — may reach:
- Sandboxed (default): anything it makes is written to a review folder for you to accept or discard. It cannot reach your project folders or the internet.
- Full access: it works directly in your real project folders and can call outside services, unattended. Switching to Full shows a confirmation dialog first — this is the only control in the app that can widen the Operator's reach.
Everything else about the Operator (turning it on, objectives, budget, the journal and review queue) lives in the Operator panel on the workspace rail.
Hooks
User-configurable lifecycle hooks: run an action when something happens during a run (an agent starts, a tool is called, a wave completes, …). No-code actions — block-if-matches, notify, webhook — plus a script/HTTP escape hatch for power users. A master kill-switch and per-hook enable toggles control what actually fires. Paid feature: on the free tier you can still edit hooks, they just don't run.
Learning
The layer of onboarding scaffolding, contextual guidance, and sample content.
| Field | What it does |
|---|---|
| Master toggle | When off, the entire learning layer goes dark — no welcome modal, no guidance cards, no sample folder seeding. |
| Replay welcome | Make the welcome intro reappear on next launch. |
| Welcome slides | Edit the welcome intro slides in your own voice (up to 5 slides). |
| Template samples | The sample files seeded into your chosen project folder when you load a starter template. Editable per template. |
| Guidance cards | The contextual hint cards. Edit copy or reset to defaults. |
MCP
Manage Model Context Protocol servers. Each configured server is started at workflow-run start and its tools are merged into every agent's tool pool.
Each server row shows its status, with enable/disable, Trust/Untrust, test, and delete controls. In Auto approval mode, tools from servers you haven't marked as trusted ask for approval; trusted servers run without prompting. Adding a server takes:
| Field | Required | What it does |
|---|---|---|
| Name | Yes | Display name in the UI. |
| Command | Yes | The executable to spawn (e.g. npx). |
| Args | No | Command arguments. |
| Environment variables | No | Key/value pairs passed to the spawned process. |
| Description | No | A human-readable note. |
Web Search
Configure the web_search tool's backend: Tavily, Perplexity, or Linkup.
| Field | What it does |
|---|---|
| Default search provider | Workflow-wide default backend. |
| API key (per provider) | Masked input. Stored in the OS keychain. |
| Test | Per-provider button that validates the key before you commit to it. |
Per-agent override is set in the Agent Config popup → Search provider field.
Vault
Securely store secrets your workflows need without putting them in .cerevisor files. Three scopes:
- Universal — available to every workflow.
- Per-workflow — grouped by workflow.
- Per-folder — grouped by project folder.
A read-only Linked entries section at the bottom shows provider, search, and license keys with links to the tabs that own them. Revealing a value requires a confirmation and re-masks after 5 seconds. Vault references are resolved at run time and never written to logs. Paid feature: the free tier sees an upgrade card here.
Subscription
License management. Shows your current plan — Trial (with countdown), Free, Paid, or Grace (license server temporarily unreachable) — plus the management actions: paste a new key, re-validate, deactivate this device (frees the seat), and open the customer portal.
Updates
Auto-update controls: current version, last-checked time, status, a manual Check for updates button, download progress when applicable, and a Restart now button once an update has downloaded.
Updates always work regardless of your license tier — paid customers cannot be stranded on a broken build by a license-server outage.
Reset
A selective, double-confirmed factory reset. Tick which categories of local data to erase — memories, saved workflows & drafts, run history & analytics, provider keys, Vault secrets, settings, skills caches — and Cerevisor wipes only those, then relaunches. Your activated license is always kept and is never an option to erase.