Intelligence Assistant

Intelligence Assistant for Obsidian

Transform your vault into an AI-native workspace. Intelligence Assistant brings a fast chat interface, configurable agents, CLI agent SDKs, MCP tool integrations, RAG over your notes, web search, and HTTP/OpenAPI tools — all without leaving Obsidian.

Desktop only. Requires Obsidian v0.15.0+.

Feature Highlights

• Modern Chat View — Streaming chat with provider/model badges, token usage display, tool-call visualization, and an Agent mode that runs multi-step agentic loops.
• Multiple LLM Providers — OpenAI, Anthropic, Google Gemini, DeepSeek, Ollama (local), OpenRouter, and SAP AI Core. Add your own credentials per provider.
• Configurable Agents — Define reusable agents with custom prompts, tool permissions, memory settings, and MCP server access.
• CLI Agent SDK Bridge — Run Claude Code, Codex, and Qwen Code as full SDK-backed agents. Each CLI tool loads its own project config (CLAUDE.md, AGENTS.md, QWEN.md) natively; the plugin only configures what the tools can't handle themselves.
• MCP Integration — Connect any Model Context Protocol server. Tool catalogs are cached for instant reuse. Includes a built-in MCP inspector for live testing.
• RAG — Index your vault with a local vector store and inject relevant notes as context on every query.
• Web Search — DuckDuckGo, Google CSE, Bing, Tavily, Brave, SerpAPI, SearXNG, Qwant, Mojeek, Yahoo, and Yandex. Configurable locale, freshness, and allow/deny lists.
• HTTP / OpenAPI Tools — Point the plugin at any OpenAPI spec (local or remote); every path/verb pair becomes an agent tool automatically.
• Quick Actions — Pre-configured shortcut buttons for common tasks (summarize, explain, translate, etc.).
• Context Attachments — Attach files, images, or vault references (@mentions) to any message.

Quick Start

Community Plugin (recommended)

1. Open Settings → Community plugins in Obsidian.
2. Browse for "Intelligence Assistant" and install.
3. Enable the plugin, then open Settings → Intelligence Assistant to configure providers.

Manual Install

1. Download the latest release.
2. Extract into <vault>/.obsidian/plugins/intelligence-assistant/.
3. Reload Obsidian and enable the plugin.

Requirements

Chat Experience

1. Open the chat via the ribbon icon or Command Palette → Open Intelligence Assistant.
2. Choose Chat or Agent mode in the toolbar. Agent mode runs the model in a loop with tool calls.
3. Pick a provider, model, temperature, and token cap from the header controls.
4. Toggle RAG (vault context) or Web search in the input bar.
5. Attach files or vault references with @filename.
6. Every response shows the provider, model, and token usage. Use Insert to Notes to drop a reply into the active file.

Agents

Regular Agents

Defined under Settings → Agents. Each agent has:

• System prompt, tool permissions, memory behavior
• MCP server access (optionally scoped)
• Custom icon and display name

CLI Agents

Defined under Settings → CLI Agents. Each CLI agent uses an SDK bridge to run Claude Code, Codex, or Qwen Code as a subprocess. Configuration follows a thin-bridge approach:

Enable Use project settings (default on) to let each CLI tool load its own project config files automatically.

SDKs are installed on-demand via the Install SDK button in the detection panel.

MCP Tools

Connect MCP servers under Settings → MCP. The plugin:

• Connects at startup and caches the tool manifest
• Injects MCP tools into agent loops automatically
• Provides a live MCP Inspector for testing tools interactively
• Supports stdio and SSE transports

Web Search

Configure under Settings → Web Search. Pick one of 11 supported providers and set locale (language + region), result freshness, and domain filters. Credentials are stored per vault.

HTTP / OpenAPI Tools

Add OpenAPI specs under Settings → Tools → HTTP / OpenAPI. For each source:

• Point at a local .json/.yaml file or a remote URL (cached locally)
• Override the base server URL and inject auth headers or query params
• Enable/disable the source; reload tool definitions on demand

Every endpoint becomes an agent tool without writing any code.

Development

npm install          # install dependencies
npm run dev          # development build + file watcher
npm run lint         # ESLint (src/ + main.ts)
npm run type-check   # TypeScript type check (no emit)
npm run test         # Jest test suite
npm run build        # production bundle

Additional scripts: npm run dev:hot, npm run build:analyze, npm run deploy:local.

Post-task rule

After every change, run npm run lint && npm run build. Both must succeed before the work is considered done.

Documentation

Contributions, issues, and feature requests are welcome — open a PR or discussion.