AgentBoard

A switchboard for AI in your browser: wire in any model, script WebMCP tools, connect remote MCP servers, bring your commands.

• Multi agent: Configure as many profile as you want, switch mid-conversation.
• Your provider: OpenAI, Anthropic, Google, or your own completion-compatible endpoint.
• Your settings: System prompts, temperature, thinking settings.
• Your keys: Bring your own API keys. No lock-in, no upselling.
• Your tools: Script WebMCP tools for page interactions. Connect remote MCP servers.
• Your commands: Template prompts with arguments; type /analyze, not paragraphs.

Bring your own models—local, fine-tuned, custom—to power multiple agent profiles. Connect remote MCP servers: bring external tools (APIs, databases, services) into browser context. Script WebMCP tools that interact with page content — think, Greasemonkey for the AI age. Built for power users.

Install extension from Chrome Web Store

Architecture overview

  ┌────────────────────────────────────┐
  │  === Tab-Scoped AI Sidebar ===     │
  │  - Markdown rendering              │
  │  - Tool visualization              │
  │  - Reasoning display               │
  └────────────┬───────────────────────┘
               │ Port (streaming)
  ┌────────────▼───────────────────────┐
  │  Background Service Worker         │
  │  ┌──────────────────────────────┐  │
  │  │ AI Client                    │  │
  │  │  - Multi-provider            │  │      ┌──────────────────────────┐
  │  │  - Streaming + tools         │  │      │      Browser Tab         │
  │  └──────────────────────────────┘  │      │  ┌────────────────────┐  │
  │  ┌──────────────────────────────┐  │      │  │ ISOLATED: Relay    │  │
  │  │ ToolRegistry                 │  │◄─────┤  │  - Forwarding      │  │
  │  │  - WebMCP + Remote + System  │  │ Port │  └────────┬───────────┘  │
  │  └──────────────────────────────┘  │      │           │              │
  │  ┌──────────────────────────────┐  │      │  ┌────────▼───────────┐  │
  │  │ TabManager                   │  │      │  │ MAIN: window.agent │  │
  │  │  - Script injection          │  │      │  │  - Tool execution  │  │
  │  │  - Lifecycle                 │  │      │  │  - Full DOM access │  │
  │  └──────────────────────────────┘  │      │  └────────────────────┘  │
  └────────────────────────────────────┘      └──────────────────────────┘

The AI sidebar is tab-scoped—each sidebar instance binds to one browser tab and sees tools from that tab's page context plus global tools (remote MCP servers and system capabilities). When you open the sidebar, the background service worker injects WebMCP scripts into the page: a polyfill that provides window.agent, a relay in the ISOLATED world, and a bridge in the MAIN world. Pages register tools via window.agent.registerTool(), triggering tools/listChanged events that flow through a persistent port connection from the relay to the background service worker. The ToolRegistry aggregates all discovered tools—WebMCP from pages, remote MCP from external servers, system tools from the extension—and converts them to AI SDK format for the AI Client.

When you send a message, the AI Client streams responses from your chosen provider (OpenAI, Anthropic, Google, or custom endpoint). Tool calls route based on type: WebMCP tools execute in the browser tab's MAIN world via JSON-RPC callTool commands (full DOM access, zero serialization overhead), remote MCP tools execute on your external servers with streaming HTTP, and system tools run in

... View full README on GitHub