✦ What is Thought Agent?

Thought Agent is not a chatbot bolted onto your notes. It is an autonomous reasoning agent that navigates your vault like a researcher, searching semantically, traversing graph links, proposing safe edits, and generating visual diagrams, all without touching a single file until you approve.

"Think of it as a second brain for your second brain."

📦 Installation

Option 1, Obsidian Community Plugins (recommended)

1. Open Obsidian → Settings → Community Plugins → Browse.
2. Search for "Thought Agent".
3. Click Install, then Enable.

That's it — no BRAT, no manual file copying.

You can also open the plugin page directly: community.obsidian.md/plugins/thought-agent

Option 2, BRAT

BRAT lets you install plugins directly from GitHub and get pre-release updates.

1. Install Obsidian42 - BRAT from the Community Plugins store.
2. Open Obsidian → Settings → BRAT → Add Beta Plugin.
3. Paste the repository URL:

   https://github.com/Tuguberk/Thought-Agent-Obsidian
   

4. Click Add Plugin, BRAT downloads and enables it automatically.
5. Go to Settings → Community Plugins and toggle Thought Agent on.

Option 3, Manual installation

1. Go to the Releases page and download the latest:
   • main.js
   • manifest.json
   • styles.css
2. Copy all three files into your vault's plugins folder:

   <your-vault>/.obsidian/plugins/thought-agent/
   

3. Restart Obsidian (or reload without saving).
4. Go to Settings → Community Plugins and toggle Thought Agent on.

Option 4, Build from source

# Clone into your vault's plugins folder
git clone https://github.com/Tuguberk/Thought-Agent-Obsidian \
  "<your-vault>/.obsidian/plugins/thought-agent"

cd "<your-vault>/.obsidian/plugins/thought-agent"
npm install
npm run build

Then enable the plugin in Obsidian as above.

🚀 Quick Start

Step 1, Configure a provider

Open Obsidian → Settings → Thought Agent and pick your LLM provider:

Using Anthropic (Claude):

• Set Provider → Anthropic
• Paste your Anthropic API key (sk-ant-...)
• Select a model, Claude Sonnet 4.6 is recommended for the best balance of speed and power

Using a local model (OpenAI compatible API):

• Start any OpenAI-compatible local server (e.g. LM Studio, Ollama, llama.cpp)
• Set Provider → OpenAI compatible API
• Default URL is http://localhost:1234/v1, hit Test to confirm it's reachable
• Your prompts never leave your machine 🔒

Step 2, Index your vault

In Settings → Thought Agent, click Re-index vault.
This builds the local semantic index (~25 MB model download on first run). You only need to do this once, and again after adding many new notes.

Step 3, Open the chat

Click the 🧠 Thought Agent icon in the left sidebar (or run "Open Thought Agent chat" from the Command Palette).

Example prompts to try:

"What do my notes say about machine learning?"
"Summarize my research on [[Attention Mechanisms]] and link related notes"
"Create a mindmap of my project ideas"
"Find notes I haven't linked yet about distributed systems"
"What are the backlinks to my MOC note?"

Step 4, Review & approve changes

The agent will never write to your vault without showing you a preview first.
Every proposed note creation, edit, and diagram goes through a diff view, approve or reject each change individually.

🖼️ Multimodal Input

Attach images and PDFs directly to your messages using the + button in the chat composer.

• Images (JPEG, PNG, GIF, WEBP) — sent as-is to vision-capable models
• PDFs — text is extracted client-side via Obsidian's built-in PDF.js and sent as readable content; no PDF data leaves your machine for local models
• Attach multiple files at once; each appears as a chip with a thumbnail preview
• Attachments are cleared automatically after each message

🏗️ System Architecture

flowchart TD
    U(["👤 User Prompt"]) --> AL["🔄 Agent Loop\nIterative reasoning"]

    AL --> TE["⚙️ Tool Executor"]

    TE --> R["🔍 Hybrid Retrieval\nVector · BM25 · Graph"]
    TE --> V["📝 Vault Operations\nCreate · Edit · Link"]
    TE --> D["🎨 Excalidraw Engine\nCreate · Update · Read"]
    TE --> G["🕸️ Graph Queries\nNeighbors · Backlinks · Subgraphs"]

    R --> CP["📦 Context Pack\nRanked, deduplicated"]
    CP --> AL

    AL --> PR["👁️ Preview & Diff\nPropose before write"]
    PR --> W(["✅ Write to Vault\nUser-approved only"])

    style U fill:#7c3aed,color:#fff,stroke:none
    style W fill:#059669,color:#fff,stroke:none
    style PR fill:#d97706,color:#fff,stroke:none
    style CP fill:#1e40af,color:#fff,stroke:none

🔌 LLM Provider Support

Thought Agent is provider-agnostic. Choose the model that fits your workflow.

flowchart LR
    TA["🧠 Thought Agent\nProvider Layer"] --> AN
    TA --> LM
    TA --> OA

    subgraph AN["☁️ Anthropic (Cloud)"]
        C1["Claude Sonnet 4.6\n✦ Recommended"]
        C2["Claude Opus 4.7\n✦ Most powerful"]
        C3["Claude Haiku 4.5\n✦ Fastest"]
    end

    subgraph LM["🖥️ OpenAI Compatible API (Local)"]
        L1["LM Studio · Ollama · llama.cpp\nQwen · Mistral · Llama · Phi\nAny OpenAI-compat model"]
    end

    subgraph OA["🔗 OpenAI-Compatible (Remote)"]
        O1["Any server exposing\n/v1/chat/completions"]
    end

    style TA fill:#7c3aed,color:#fff,stroke:none
    style AN fill:#1a1a2e,color:#e2e8f0,stroke:#7c3aed,stroke-width:1px
    style LM fill:#1a1a2e,color:#e2e8f0,stroke:#059669,stroke-width:1px
    style OA fill:#1a1a2e,color:#e2e8f0,stroke:#3178c6,stroke-width:1px

🔍 Retrieval Pipeline

Every query runs through a multi-stage hybrid search before the agent sees any context.

flowchart LR
    Q["🔎 Query"] --> V["Vector Search\n(all-MiniLM-L6-v2)\nSemantic similarity"]
    Q --> B["BM25 Scoring\nKeyword frequency\n& TF-IDF"]

    V --> F["🔀 Fusion\nMin-max normalize + weighted merge\n(0.7 semantic + 0.3 BM25)"]
    B --> F

    F --> GR["🕸️ Graph Re-ranking\nBoost linked neighbors\nof top results"]
    GR --> MM["📊 MMR Filtering\nMaximal Marginal\nRelevance, diversity"]
    MM --> CTX["📦 Context Pack\nTop-K unique chunks\nwith note metadata"]

    style Q fill:#7c3aed,color:#fff,stroke:none
    style CTX fill:#059669,color:#fff,stroke:none

✏️ Safe Write Workflow

"Nothing touches your vault without your eyes on it first."

sequenceDiagram
    actor User
    participant Agent
    participant Preview
    participant Vault

    User->>Agent: "Summarize my research and link related notes"
    Agent->>Agent: Search + reason (tool loop)
    Agent->>Preview: Propose changes (diff view)
    Preview-->>User: Show before/after for each change
    User->>Preview: Approve ✅ / Reject ❌ per item
    Preview->>Vault: Write only approved changes
    Vault-->>User: Notes updated safely

🎨 Excalidraw Integration

Generate, update, and search visual diagrams, directly from natural language.

flowchart TD
    NL(["💬 'Create a mindmap\nof my ML notes'"]) --> EA["Excalidraw\nAgent Tools"]

    EA --> T1["📐 create_diagram\nmindmap · flowchart\ntimeline · entity-graph"]
    EA --> T2["✏️ update_diagram\nAdd nodes · edges\nUpdate labels"]
    EA --> T3["🔍 search_diagrams\nSemantic index\nover .excalidraw files"]
    EA --> T4["📖 read_diagram\nExtract nodes, edges\n& text content"]
    EA --> T5["🔗 annotate_diagram\nBidirectional link\nnote ↔ diagram"]

    T1 --> LE["🏗️ Layout Engine\nDeterministic placement\nAnchor-aware arrows\nLayer ordering"]
    T2 --> LE
    LE --> PR["👁️ Preview\n→ User Approval"]
    PR --> OUT(["📄 .excalidraw file\nwritten to vault"])

    style NL fill:#7c3aed,color:#fff,stroke:none
    style OUT fill:#059669,color:#fff,stroke:none
    style LE fill:#d97706,color:#fff,stroke:none

Supported diagram types:

🛠️ Agent Tool Inventory

The agent has 14 tools across two categories:

⚡ Getting Started to Development

Requirements

• Obsidian Desktop 1.7.2+
• Node.js 18+
• npm
• (Optional) Excalidraw Obsidian plugin for diagram features

Installation (Development)

# 1. Clone into your vault's plugins folder
git clone https://github.com/Tuguberk/Thought-Agent-Obsidian \
  /path/to/vault/.obsidian/plugins/thought-agent

# 2. Install dependencies
cd thought-agent && npm install

# 3. Start the dev build (watches for changes)
npm run dev

# 4. Enable "Thought Agent" in Obsidian → Settings → Community Plugins

Production build

npm run build

📄 License

MIT © Tuğberk Akbulut