1. The Model Organizer
The Model Organizer is the single place you connect models. Equipped models appear in every model picker across chat, agents, and the standalone tools.
| Provider type | Notes |
|---|---|
| Google Gemini | Paste a Gemini API key. Keys are validated live before being saved — a key is not accepted just because it starts with AIzaSy. |
| OpenAI / Anthropic / Groq / DeepSeek / xAI / etc. | Built-in catalog entries; enter the provider key. Proxy/fallback logic mirrors each provider's native API. |
| Custom OpenAI-compatible | Any endpoint speaking the OpenAI schema (OpenRouter, Together, local servers, Azure-style). Use Discover models to auto-list /models; per-model API keys are supported. Remote endpoints go via the allowlisted proxy; local URLs are called directly. |
| Custom Anthropic-compatible | Anthropic Messages-API-compatible endpoints, with the anthropic-dangerous-direct-browser-access header and proxy-unreachable direct fallback. |
| Ollama (local) | Set your Ollama URL (default http://127.0.0.1:11434) and launch with OLLAMA_ORIGINS="*" ollama serve so the browser may call it. Connection status is polled and shown inline. |
| WebLLM / Transformers.js | Open-weights models run in-browser on WebGPU/WASM. Load once; weights are cached. Switching models unloads the previous engine to free GPU/WASM memory. |
| Gemini Nano | Chrome's on-device model; auto-appears in the catalog when supported. |
| Puter.js (free) | Free, proxy-based inference with no key; useful to start immediately (see Part 2). |
2. Provider Test Arena
The organizer can run a live prompt against multiple equipped models concurrently per provider and sequentially within a provider (so local single-engine models do not contend), reporting first-token latency and tokens-per-second. Use it to compare real performance before committing a default.
3. Tool Permission Levels
Tool calls can originate in model output, imported files, or MCP responses, so a synchronous confirmation sits at the final dispatch point and cannot be bypassed by an alternate agent loop.
- Permission levels control which tools run automatically, which ask first, and which are blocked. Write-class tools (file writes, commits, external actions) are classified separately from read-only ones.
- Built-in vs external. The built-in web-search and GitHub connector tools are distinguished from external MCP servers by server identity; an external server cannot impersonate a built-in tool by naming it
github_*. - AI memory and skill directives. Creation proposals require the corresponding
aiCanCreateSkillsoraiCanCreateMemoriessetting and a separate Keep click in the chat review card; proposals expire after 60 seconds and are capped at two per conversation. Deletions requireaiCanDeleteSkillsoraiCanDeleteMemories. All four settings default off.
4. Connectors (GitHub & Google Drive)
GitHub: connect via OAuth (recommended) or paste a Personal Access Token with repo, read:user, user:email scopes. Disconnect fully clears tokens and refresh credentials. You can also open a local folder in read-only mode to browse code without pushing.
Google Drive: used for optional backups of chats, memories, projects, files, and settings (secrets stripped). The access token is session-scoped; sync selectors let you choose exactly which slices back up.
5. Generation & Display Settings
Temperature, max tokens, context window limit, Infinity Mode (repeated self-correction attempts + critic strictness), auto-architecture routing, and vision toggle.
Image-generation behavior and error handling. The current Settings panel does not expose model or provider selection controls; an error card may offer a manual Pollinations retry.
Drive sync selectors (chats/memories/projects/files/settings/skills), export/import snapshots, and local-folder sandbox quotas.
Theme, Robot Companion visibility, developer-density options, and the EULA/terms acknowledgement.
6. Where Settings Persist
Settings save to IndexedDB and are mirrored to localStorage for fast bootstrapping. API keys are isolated in the credential store; Drive sync stores them only in a separate AES-GCM encrypted vault and its private AppData key record, never in the readable settings document. Signing into the same Google account restores that vault on another device. Changing a setting updates persistence without double-writing, so concurrent edits (such as a background Drive sync) are not clobbered.
7. The Settings Search Bar — a Field Guide
The modal's instant search is the fastest route to any toggle; these are the high-traffic queries and where they land:
ollamaEngine & API → local endpoint URL and connection status
puterModel Organizer → free Puter.js provider and custom model IDs
fallbackModel Fallbacks — saved order (not applied at request time, Part 8)
temperatureLLM Parameter Engine — sampling controls
infinityMulti-Agent Systems — critic strictness & max attempts
queueAI Engine — prompt queue backoff interval (0–60s)
aiCanCreateAI Permissions — skill/memory creation & deletion gates (default off)
syncGoogle Drive Cloud — per-slice sync selectors
purgeLocal Backup & Export — irreversible workspace wipe
Search matches labels and keywords — if a query comes up empty, try a synonym ("clear" for purge, "delete" for remove) before hunting through tabs.
8. Three Reference Configurations to Steal
Default AL Router Fast + Puter models; Drive backup off; AI create/delete permissions all off; Auto Router on; 25–30 new chats before evaluating anything. Spend nothing, learn the product, upgrade deliberately.
Groq direct key for speed, one OpenRouter key for depth, Ollama for confidential code; Drive sync on for chats+skills+memories with a monthly restore drill (Part 17); AI skill creation on with Keep-review habit; Multi for builds, Single for everything else; fallback order saved but switched manually on failure; weekly snapshot export.
WebLLM / Transformers.js / Gemini Nano / local Ollama only; cloud providers disabled or keyed to a personal paid account; memories off or minimal; Drive sync off — snapshot exports to encrypted local disk; all AI permissions off; sandbox-first workflows; clearing routine after each session on shared networks.
9. The Provider Test Arena — Running It and Reading It Right
Numbers that look comparable are not always comparable. A sound protocol:
- Use one fixed canary prompt — the same task shape across all your equipped models (e.g. "Explain CRDTs to a senior Go engineer in 250 words") so differences come from the model, not the ask.
- Mind the execution pattern. Providers run concurrently across providers but sequentially within a provider — local engines and same-vendor models don't contend, but your timing conclusions should come from separate providers on separate runs, never from one shared benchmark morning.
- Read TPS with its denominator. Tokens/sec on a 40-token answer vs a 900-token code task are different populations — compare like with like, or use first-token latency for the "does it feel alive" judgment.
- Latency vs. quality are different axes. The arena will rank a tiny fast model above a slower strong one on speed — pick your default per task class, not per leaderboard.
- Re-run after provider events. Degradation, new model drops, and quota resets all scramble yesterday's ranking.
10. A Concrete Walkthrough: The Write-Tool Ask-First Flow
What actually happens when an agent wants to write through the GitHub connector mid-task:
- The model emits a tool call for
github_create_or_update_file(or similar write-class built-in). - The dispatch gate intercepts it. Write-class classification happens before execution; the call never reaches GitHub yet. Read-class built-ins in the same run proceed unimpeded.
- A confirmation card appears above the composer summarizing the target repo, path, and action. This is not a toast you can fat-finger past — it blocks until you decide.
- Approve → the call executes and the result streams back into the run; deny → the model receives a refusal it can react to (usually: propose a plan instead, or ask you for explicit instruction); ignore → the card expires and the tool never runs.
- Audit afterward: the run's activity trace records the attempt and its resolution — review it when you let an unattended batch touch anything write-class.
Defaults exist for a reason: creation/deletion of skills and memories, plus all write-class tool calls, start ungranted. Grant narrowly, watch the first few executions, then decide whether to keep the ask-first friction.