Transformers.js for Chrome Extensions

Outlines an architecture for attaching Transformers.js to a Chrome extension under Manifest V3 constraints.

• The background service worker handles model hosting, agent state, and tool execution.
• The side panel handles the chat UI and streaming responses.
• The content script handles page extraction and highlighting.

Models are split into two roles. Gemma 4 E2B ONNX handles text generation and tool selection, while all-MiniLM-L6-v2 ONNX handles embeddings for search. All inference runs in the background, and the cache is shared and stored under the chrome-extension:// origin to reduce redundant loading. Since the service worker can be suspended, model state must be handled in a re-initializable way.

The execution loop works by parsing tool_call from model output and converting it into deterministic tool execution. webMcp normalizes extension tools into a name / description / inputSchema / execute format, and extractToolCalls separates messages from tool calls. Example tools include tab control, opening URLs, history search, querying web pages, and highlighting elements.

State is separated by purpose.

• Conversation history: background memory
• Settings: chrome.storage.local
• Vector history: IndexedDB
• Page content: background cache

The key point is that local AI extensions need a background-executor-centric structure, a strict messaging contract, and a model caching strategy.