🤖 AGI Desktop Intelligence Agent

A privacy-first desktop assistant that uses a local LLM (Ollama), Linkup's agentic search API for real-time web knowledge, and 12 tools across 5 domains — all orchestrated by an autonomous Planner → Executor → Evaluator loop.

🔗 Interactive Architecture Diagram — animated, clickable system overview (or open docs/architecture_diagram.html locally)

---

Features

Multi-Domain Intelligence — Email management, document analysis (PDF/DOCX/TXT), calendar scheduling, web research, and semantic memory — all through natural language.

Privacy-First — PII is automatically redacted before the LLM sees it. Names become <PERSON_1>, phones become <PHONE_1>. Company names, dates, and task-relevant context are preserved. Real values are restored in the final response and generated files.

Linkup Web Search — Agentic search for real-time research, fact-checking, and meeting prep. The agent decides when external knowledge is needed — not on every prompt. Only sanitized queries are sent.

Autonomous Planning — The agent decomposes complex multi-step requests into executable plans, selects tools, evaluates results, and retries on failure. A scheduling gate enforces that calendar events are only created when the user provides a confirmed date/time.

Generated Files — Email drafts open directly in your mail app (Outlook, Gmail, etc.) via mailto:. Calendar events download as .ics files. Both are accessible from download buttons in the chat.

---

Screenshots

  

Setup screen checks all prerequisites on launch

Multi-step workflow: email draft + scheduling gate + calendar event creation with privacy redaction (AI View panel)

---

Quick Start

Prerequisites

• Python 3.10+
• Ollama — install and make sure the Ollama service is running (ollama serve)
• Linkup API key — sign up free at linkup.so → Dashboard → API Keys

Setup

# 1. Clone the repo
git clone https://github.com/GitReboot/HackwithDC-Team-6.git
cd HackwithDC-Team-6

# 2. Install Python dependencies
pip install -r requirements.txt

# 3. Pull an Ollama model (make sure Ollama is running first)
ollama pull llama3.1:8b

# 4. Set your Linkup API key
# Windows CMD:
set LINKUP_API_KEY=your-key-here
# Windows PowerShell:
$env:LINKUP_API_KEY="your-key-here"
# Mac/Linux:
export LINKUP_API_KEY="your-key-here"

# 5. Start the web server
python server.py

# 6. Open http://localhost:5000 in your browser
#    The setup screen will check all prerequisites automatically

Demo emails and calendar events auto-seed on first launch — no extra setup needed.

Optional: Better PII Detection

pip install presidio-analyzer presidio-anonymizer spacy
python -m spacy download en_core_web_sm

Without this, privacy still works using regex patterns.

Recommended Models

Model	RAM	Tool Calling	Best For
llama3.1:8b (default)	8 GB	Good	Any system
qwen2.5:7b	8 GB	Very Good	Better structured output
qwen2.5:14b	12 GB	Excellent	Best results if hardware allows

Change model in config/defaults.yaml:

agent:
  model: "qwen2.5:7b"

---

Architecture

User Input → Privacy Layer → Planner → Executor → Evaluator → Synthesis → Response
                                          │
                    ┌─────────────────────┼──────────────────────┐
                    │                     │                      │
              Linkup Search      Local Tools (12)         Memory Store
              (web research)     Email · Docs · Calendar   SQLite + FAISS

The agent follows a Planner → Executor → Evaluator loop:

1. Planner decomposes goals into steps
2. Executor calls tools via LLM function-calling (with scheduling gate + PII restore)
3. Evaluator checks success and retries if needed
4. Synthesis combines results, strips false claims, restores PII

See docs/architecture.md for the full design, or explore the interactive diagram.

---

Project Structure

desktop-agent/
├── main.py                    # CLI entry point
├── server.py                  # Flask web server + API
├── requirements.txt
├── config/
│   ├── __init__.py            # Config loader
│   └── defaults.yaml          # Default settings
├── agent/
│   ├── __init__.py
│   ├── planner.py             # Goal → step decomposition
│   ├── executor.py            # Tool calling + scheduling gate
│   ├── evaluator.py           # Success checking + retry
│   ├── loop.py                # Main orchestration loop
│   └── prompts.py             # System/plan/eval prompts
├── tools/
│   ├── __init__.py            # BaseTool + ToolRegistry
│   ├── linkup_client.py       # Linkup agentic search
│   ├── email_adapter.py       # IMAP + local mailbox + mailto
│   ├── document_adapter.py    # PDF/DOCX/TXT reader
│   ├── calendar_adapter.py    # .ics calendar (zero-dep)
│   ├── memory_tools.py        # Store/recall facts
│   └── privacy.py             # PII redaction engine
├── memory/
│   ├── __init__.py
│   ├── sqlite_store.py        # Conversation + task history
│   └── faiss_retriever.py     # Semantic similarity search
├── frontend/
│   └── index.html             # Web chat UI
└── docs/
    ├── architecture.md         # System design document
    ├── architecture_diagram.html  # Interactive animated diagram
    ├── linkup_integration.md   # Linkup usage documentation
    └── demo_scenarios.md       # Demo walkthrough

---

Documentation

Document	Description
architecture.md	Full system design — data flow, components, guardrails
architecture_diagram.html	Interactive animated architecture diagram
linkup_integration.md	How Linkup enhances agent capabilities
demo_scenarios.md	Step-by-step demo walkthrough

---

Evaluation Criteria Mapping

Criteria	Implementation
Generality	12 tools across 5 domains — email, documents, calendar, web search, memory
Autonomy	Planner → Executor → Evaluator loop with minimal hardcoded flows
Reasoning	LLM-driven planning, multi-step tool chaining, scheduling gate
Context Awareness	Short-term buffer + long-term semantic memory (FAISS)
Information Synthesis	Linkup for real-time web knowledge, result grounding
Privacy & Security	PII redaction, local-first processing, no data leaves machine
Usability	Web UI with chat, file attachments, AI View, generated file downloads

---

Health Check

Visit http://localhost:5000/api/health after starting the server to verify all prerequisites:

• Python version
• Ollama status and model availability
• Linkup API key
• Required packages
• Privacy engine (Presidio NLP or regex fallback)
• Demo data

The web UI also shows an interactive setup screen on launch that runs through these checks automatically.

---

Future Scope

• Streaming responses — Real-time token-by-token output instead of waiting for the full response, significantly improving perceived speed
• Conversation history — Sidebar showing past sessions with the ability to resume or reference previous conversations
• Multi-agent architecture — Separate planner, researcher, and critic agents that collaborate on complex tasks
• OAuth integrations — Direct Google Calendar, Outlook, and Gmail connections instead of local file adapters
• Voice input — Speak commands instead of typing, with local speech-to-text processing
• Plugin system — Allow users to add custom tools without modifying core code

---

Powered By

  

Ollama powers local LLM inference  ·  Linkup provides agentic web search