Desktop only · Models, tools and vault context in one workspace.

From a request to saved notes: tool use, verification and a visible task record.

✨ Features

• 💬 Chat and Agents: streaming responses, note/folder references, model-dependent image input, reusable prompts and Agent presets: Librarian, Architect, Researcher and Scribe.
• 🧰 Vault tools: 13 built-in tools for reading, listing and searching files; creating, updating and appending notes; properties; Canvas; PDFs; tags and links. See the tool reference. Built-in move/delete tools are not included.
• 📝 Write proposals: built-in Agent writes normally produce reviewable proposals. An approved task can apply them during execution when the Agent's Autonomous write setting is enabled.
• ⏯️ Manual continuation: use Stop, then send Continue. in the same conversation, including after reloading Obsidian. Saved task context and tool evidence support continuation; this is not automatic crash replay or an exactly-once guarantee for external actions.
• 📚 Knowledge and memory: optional RAG with a local index, a configurable embedding provider, source references, and per-Agent memory modes: none, short-term or long-term.
• 🔌 Integrations: MCP stdio servers, OpenAPI HTTP tools and local CLI tools. Web search has eight configurable backends; credentials, endpoints and availability depend on the backend.
• 🌐 Models: OpenAI, Anthropic, Google Gemini, DeepSeek, OpenRouter, Ollama, SAP AI Core and custom OpenAI-compatible endpoints. Tool calling and image support depend on the selected model and service.
• 🎨 Interface: nine top-level settings Tabs, light/dark themes and 49 bundled UI locales.

🖼️ A look inside

Choose the tools your Agent can use	Configure retrieval for your vault

Actual Obsidian captures. Screenshot provenance and validation scope.

📦 Installation

Community plugins

1. Open Settings → Community plugins → Browse.
2. Find Intelligence Assistant, install it and enable it.
3. Check Settings → Intelligence Assistant → General for the installed version.

The plugin is listed in the official community registry. Community installation uses published release assets; it may install a different version from this checkout. See Releases for published assets, the 0.1.1 notes for cumulative changes including 0.1.0, and the 0.1.0 record for its original validation scope.

From source or release files

To build this checkout, use Node.js 22 and the committed lockfile:

git clone https://github.com/qwai-tech/obsidian-plugin-intelligence-assistant.git
cd obsidian-plugin-intelligence-assistant
npm ci
npm run build:production

Copy main.js, styles.css and manifest.json into <vault-config>/plugins/intelligence-assistant/ (<vault-config> is usually .obsidian inside your vault). Reload Obsidian and enable the plugin. For an upgrade, back up your vault and replace those three application files; retain data.json and the plugin's data/ folder.

🚀 First setup

1. Open Settings → Intelligence Assistant → LLM → Providers. Add the provider and select the appropriate credential from Obsidian's Keychain. Custom endpoints may also need a base URL.
2. Test the connection, review LLM → Models, enable an available chat model and set a default. A reachable endpoint with no usable models does not establish that generation works.
3. Open AI chat from the ribbon or command palette. Start with a short chat request.
4. For tool use, select Agent mode and review the Agent's model strategy, tools, step/output budgets and write setting under Agents.
5. Configure RAG, memory, web search and external tools only when needed. See the user guide.

SAP AI Core uses a Keychain entry containing the complete Service Key JSON and a resource group (usually default). The AI API URL comes from serviceurls.AI_API_URL unless overridden. Connection testing discovers running deployments. SAP code is included, but this preparation did not perform a live SAP tenant test.

🔐 Permissions and data

Built-in Agent proposals govern supported vault edits. Configured CLI commands, MCP servers and HTTP operations can have their own filesystem, process and network effects; they are not sandboxed by the proposal UI. Grant an Agent only the tools it needs. Prompt text such as “without confirmation” does not by itself enable automatic vault writes.

Other enabled plugins share Obsidian's host environment. Tool permissions and secret references do not provide process isolation. See data and execution details.

⏱️ Long tasks

Agent completion text is not independent proof that every requested artifact exists. Check the actual files and execution trace. Step/output budgets and provider timeouts still apply; interrupted external operations may need inspection before retrying.

Use exact tool names for a task restriction, such as Use only read_file; or No tools.. These narrow configured permissions and survive a bare continuation. They do not grant new tools. The accepted directive grammar is intentionally limited; see the guide.

The recorded 0.0.23 real gate completed a 40.525-minute task with 100 articles and a real file-measurement tool. The 0.1.0 version revision received build and actual installation/reload checks. These are separate scopes, documented in 0.1.0 validation.

🛠️ Development

npm ci
npm run dev                # development build and watch
npm run lint
npm run typecheck
npm run build              # development bundle
npm run build:production   # production bundle and provenance

npm run verify, Jest, missions and the CI E2E suite include mocks. For a no-mock release check, configure real services and use npm run verify:release:real; see testing. A normal push to main/master triggers Build and Publish: check the version tag, lint/type-check/build, verify artifact hashes, then create the Release automatically. The CI real-verification job is temporarily removed by explicit approval; the local real-test runner remains available. See the release process.

📖 Documentation

Issues and feature requests are welcome. Include the plugin/Obsidian version, provider/model, steps to reproduce and redacted logs.