Configuration

Environment Variables

Variable	Default	Description
PORT	8000	Preferred starting port for the backend. If busy, scans 8001–8099.
SILICON_AUTH_TOKEN	(unset)	Shared secret between Electron and backend. When set, all non-public API endpoints require Authorization: Bearer <token> (or ?token= for SSE). Unset = no auth (dev mode).
SILICON_PARENT_PID	(unset)	PID of the Electron parent. Backend polls every 5 s and exits if the parent is gone. Set automatically when launched by Electron.

The backend binds to 127.0.0.1 only. No external network access.

Frontend Settings

Settings are stored in the browser's localStorage and persist across sessions.

Chat Settings

Key: silicon-studio-chat-settings

Field	Type	Default	Description
systemPrompt	string	"You are a helpful AI assistant running locally on Apple Silicon."	System message prepended to every conversation
temperature	number	0.7	Sampling temperature (0 = deterministic, 2 = max randomness)
maxTokens	number	2048	Maximum tokens per response
maxContext	number	4096	Conversation context window size
topP	number	0.9	Nucleus sampling threshold
repetitionPenalty	number	1.1	Penalty for repeated tokens
reasoningMode	string	"auto"	One of: off, auto, low, high
webSearchEnabled	boolean	false	Inject web search results into context

These values can be changed in the Settings page or per-conversation via the collapsible parameters sidebar.

RAG Settings

Key: silicon-studio-rag-settings

Field	Type	Default	Description
chunkSize	number	512	Characters per chunk during ingestion
chunkOverlap	number	50	Overlap between consecutive chunks

Per-Conversation Overrides

The parameters sidebar in the Chat view stores overrides in component state. These are not persisted to localStorage — they reset when switching conversations. The sidebar also controls:

Setting	Description
Show Prompt	Display the raw system prompt sent to the model
Syntax Check	Auto-validate code blocks in responses
Auto-fix Syntax	Show a "Fix" button on invalid code
Memory Map	Auto-summarize context every N messages
PII Redaction	Replace emails, phones, IPs with tokens before sending
RAG Knowledge	Enable and select a knowledge collection
Web Search	Enable DuckDuckGo search injection
Visible Actions	Toggle which quick actions appear on responses

Backend Storage

All backend data lives in ~/.silicon-studio/. There is no database — everything is JSON files on disk.

models.json

{
  "model-id": {
    "id": "model-id",
    "name": "Qwen3-1.7B-MLX-8bit",
    "size": "1.7B",
    "architecture": "qwen2",
    "downloaded": true,
    "local_path": "/Users/you/.silicon-studio/models/...",
    "is_finetuned": false
  }
}

mcp_servers.json

{
  "server-uuid": {
    "id": "server-uuid",
    "name": "filesystem",
    "command": "npx",
    "args": ["-y", "@modelcontextprotocol/server-filesystem", "/tmp"],
    "env": {},
    "transport": "stdio"
  }
}

Electron Configuration

Build config is in the root package.json under the "build" key. Key settings:

Field	Value
appId	com.silicondev.app
mac.target	["dmg", "zip"]
mac.hardenedRuntime	true
mac.icon	resources/icon.icns

The backend is bundled as a PyInstaller binary in backend/dist/silicon_server and included via extraResources.