Project Info

Winner

[The Interaction Company of California] Build with Poke (Challenge 1: iPhone Air for everyone in the team Challenge 2: iPhone 17 Pro for everyone in the team Challenge 3: iPhone Air for everyone in the team Bonus: up to $100k in cash); [Browserbase] Best Web Automation with Stagehand ($1,000)

Concierge

Inspiration

So much of modern life is hyper-connected and still strangely distant. You care about the people in your life, but between work, family, and everything else, the little moments slip by; Remembering a friend’s favorite color, noticing a new hobby, picking a gift that says “I know you.” Gift-giving used to be how we showed "I was paying attention", but now it feels like another task on an impossible list. We wanted to bring back that sense of human connection in a world of disconnect. Concierge reminds you what you already know, and helps people show up for each other more thoughtfully.

What it does

Concierge is an Agentic gift-giving workflow built on top of Poke that helps you find deeply personalized gifts through peer-to-peer agent collaboration and research. When you want to buy a gift for someone, your agent talks to their agent. Your recipient’s agent represents their tastes and preferences, and runs an autonomous iMessage subagent that validates gift hypotheses by searching through the recipient’s broader message history, far more than the gSuite knowledge base Poke provides. Concierge recommendations aren’t generic; they’re grounded in real evidence. It then returns a small set of gift ideas with message citations (and surrounding context) showing exactly why each gift fits. Finally, a browser automation agent uses Browserbase + Stagehand to find the right gift, configure the right variant (size, color, model), and generate a ready-to-checkout link with an estimated final price.

How we built it

We built Concierge as a web app with a real-time dashboard (Next.js + FastMCP) that visualizes agent-to-agent negotiation and tunes the user into a gift ideating chain of thought tree. Under the hood, we extended Poke via MCP servers: a Poke-to-Poke MCP for agent communication, an iMessage Query MCP that searches the local macOS iMessage database using structured SQL queries, and a Browserbase MCP wrapper for shopping workflows. Poke B (the recipient’s agent) orchestrates the end-to-end process: generating hypotheses for , spawning the iMessage subagent loop, synthesizing validated ideas, and calling the browser agent to produce checkout links. We store session metadata and visualizations, but sensitive message data stays local.

Challenges we ran into

The biggest challenge was balancing “wow” personalization with privacy and reliability. iMessage data is messy, unstructured, and incredibly sensitive. We needed tooling that could search it quickly, extract context, and produce citations without dumping a user’s entire private life into an LLM prompt. Browser automation was also challenging, retail sites vary wildly, and checkouts can break. We had to scope carefully and focus on a flow we could make stable for a demo.

Accomplishments we're proud of

We’re proud that Concierge feels human. It doesn’t just recommend “a gift”; it shows the reasoning and evidence behind personalization, which makes it trustworthy. We also built a compelling agent visualization that demonstrates autonomous search and hypothesis refinement in real time. And we integrated real-world tools (Poke’s GSuite knowledge store + iMessage search + Browserbase checkout automation) to prove this can work at scale for anyone.

What we learned

We learned that great agent products aren’t just about model intelligence. Instead, they’re about the memory, constraints, and interfaces around the model. Evidence-backed personalization is dramatically more convincing than generic suggestions. We also learned that “agent-to-agent” is a powerful metaphor for consent, representation, and trust: the recipient’s agent can advocate for them, even in commerce.

What's next

Next, we want to expand beyond gifting into a true “preferences marketplace” where your agent can represent you across experiences, gifts, recommendations, group plans, and purchases, while keeping you in control. We’d add stronger consent flows, better redaction controls, richer preference modeling, and broader retailer support. The long-term vision is simple: use AI to help people understand each other better, and make everyday acts of care easier and more meaningful.

Concierge - AI-Powered Gift Giving

A gift-giving workflow built on top of Poke (by The Interaction Company) that finds deeply personalized gifts through peer-to-peer AI agent collaboration and autonomous message analysis.

TreeHacks 2026 Project

🎁 What is Concierge?

Concierge helps you find the perfect gift by having your Poke agent collaborate with the recipient's Poke agent to discover deeply personalized gift ideas—backed by specific message citations that prove real research.

The Problem

Generic LLMs give generic gift advice with no proof:

"Get them a keyboard, they like tech" ❌
No evidence, no personalization, no confidence

Our Solution

Concierge provides gift recommendations with citation traceability:

Recommendation: Keychron K8 mechanical keyboard (Navy, $97.19) ✅
Citation 1 (3 months ago): "I've been really into mechanical keyboards lately"
Citation 2 (2 months ago): "The clicky sound is so satisfying!"
Citation 3 (1 month ago): "My wrists hurt from this laptop keyboard"
Proof: Not generic—deeply researched with specific evidence

🚀 Key Features

1. Autonomous iMessage Analysis

An autonomous agent loop validates gift hypotheses by iteratively searching the recipient's message history:

Tests initial hypothesis
Finds contradicting evidence
Updates its own hypothesis
Continues until confident
Returns validated ideas with citations

2. Peer-to-Peer Agent Collaboration

Two Poke agents work together:

Gift Giver's Poke: Initiates request, receives recommendations
Gift Recipient's Poke: Orchestrates research, validates hypotheses

3. Full Context Access

Recipient's Poke accesses their complete message history:

Shared messages (Giver ↔ Recipient): Shared experiences
Other conversations (Recipient ↔ Others): General interests
Result: Combined context for better personalization

4. Intelligent Product Configuration

Browser automation configures products based on message insights:

Size: Extracted from messages ("I'm a size 10")
Color: From favorite color mentions
Variant: From specific preference messages

5. Real-time Visualization

Web dashboard shows the entire process:

Agent conversation in real-time
Autonomous loop iterations
Hypothesis updates (invalidated → refined → validated)
Citations as they're discovered
Final recommendations with evidence

📋 Documentation

Quick Links:

📋 PRD.md - Complete product specification (GROUND TRUTH)
🤖 agents.md - Agent architecture and collaboration patterns
📚 DOCUMENTATION_INDEX.md - Documentation hub
⚡ QUICK_REFERENCE.md - Commands & workflows

The PRD is the ground truth for all development and serves as centralized context for this project.

🏗️ Architecture

Person A (Gift Giver)
    ↓ "Buy gift for Person B, $100, tech"
┌─────────────┐
│   Poke A    │
└──────┬──────┘
       │ Poke-to-Poke MCP
       ↓
┌────────────────────────────┐
│   Poke B (Orchestrator)    │
│  - Email receipt analysis  │
│  - Hypothesis generation   │
│  - Orchestrate subagents   │
└──┬─────────────────────┬───┘
   │                     │
   ↓                     ↓
┌──────────────────┐  ┌────────────────┐
│ iMessage Subagent│  │ Browser Agent  │
│ (Autonomous Loop)│  │ (Browserbase)  │
│                  │  │                │
│ - Test hypothesis│  │ - Find product │
│ - Search messages│  │ - Configure    │
│ - Validate/Update│  │ - Get checkout │
│ - Find citations │  │   URL          │
└──────────────────┘  └────────────────┘
   │                     │
   └────────┬────────────┘
            ↓
   3 Gift Recommendations
   + Citations + URLs
            ↓
        Poke A
            ↓
      Person A

Built on: Poke by The Interaction Company
Extension Method: MCP (Model Context Protocol) servers

🚀 Quick Start

Prerequisites

Python 3.12+
Node.js 18+
uv (Python package manager)
macOS (for iMessage database access)
Supabase account
Browserbase account

Current Setup (Development)

Install Python dependencies:

uv sync

Configure environment variables:

cp env.example .env
# Edit .env with your keys (see env.example for all options):
# - POKE_A_KEY / POKE_B_KEY (agent authentication)
# - SUPABASE credentials
# - BROWSERBASE_API_KEY
# - ANTHROPIC_API_KEY or OPENAI_API_KEY

Run the Poke-to-Poke MCP server:

python mcp/poke-to-poke/server.py

📁 Project Structure

treehacks26/
├── mcp/
│   └── poke-to-poke/              # Agent communication MCP server ✅
│       └── server.py
├── agents/                         # Agent logic (to be implemented)
│   ├── poke_b_orchestrator.py     # 🔄 Orchestrates gift research
│   ├── imessage_subagent.py       # 🔄 Autonomous hypothesis validation
│   └── browser_automation.py      # 🔄 Product search & configuration
├── app/                            # Next.js dashboard (to be implemented)
│   ├── page.tsx                   # 🔄 Main dashboard
│   ├── session/[id]/              # 🔄 Gift session view
│   └── components/                # 🔄 React components
├── supabase/                       # Database migrations (to be implemented)
│   └── migrations/                # 🔄 Schema definitions
├── PRD.md                          # 📋 Product requirements (GROUND TRUTH)
├── agents.md                       # 🤖 Agent architecture reference
├── README.md                       # 🚀 This file
├── conversation.txt                # 💬 Agent communication log
├── env.example                     # 🔐 Environment template
├── pyproject.toml                  # Python dependencies
└── package.json                    # 🔄 Node.js dependencies (to be added)

Legend: ✅ Implemented | 🔄 To be implemented

🔧 Implementation Status

✅ Completed

Project structure and documentation
Comprehensive PRD (ground truth)
Poke-to-Poke MCP server (basic implementation)
Agent authentication system
Local conversation file storage

🔄 In Progress

See PRD.md Section 10 for detailed roadmap.

🧪 Testing

Test Poke-to-Poke Server

# Start server
python mcp/poke-to-poke/server.py

# Send test message
curl -X POST http://localhost:8000/mcp/v1/tools/call \
  -H "Authorization: Bearer $POKE_A_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "jsonrpc": "2.0",
    "method": "tools/call",
    "params": {
      "name": "send_message",
      "arguments": {
        "message": "[REQUEST] Test gift request"
      }
    },
    "id": 1
  }'

# Get conversation history
curl -X POST http://localhost:8000/mcp/v1/tools/call \
  -H "Content-Type: application/json" \
  -d '{
    "jsonrpc": "2.0",
    "method": "tools/call",
    "params": {
      "name": "get_conversation",
      "arguments": {}
    },
    "id": 2
  }'

Watch Conversation in Real-time

tail -f conversation.txt

🎯 Key Differentiators (Demo WOW Factor)

When demonstrating Concierge at TreeHacks:

1. Citation Traceability ⭐

Show specific messages from months ago:

Message timestamp: "5 months ago"
Full context: conversation with other friends
Proves deep research vs generic LLM

2. Autonomous Intelligence ⭐

Visualize hypothesis evolution:

Initial: "Person B likes Warriors"
Finding: 15 messages about Cavaliers instead
Updated: "Person B likes Cavaliers" ✅
Shows real reasoning, not pattern matching

3. Context Depth ⭐

Demonstrate access to full message history:

Shared experiences: Trip to Japan (from A ↔ B messages)
General interests: Loves ramen (from B ↔ C messages)
Combined insight: Japanese cooking class gift

4. Configuration Intelligence ⭐

Show attention to detail:

Found in messages: "I'm a size 10", "Navy blue is my favorite"
Product configured: Nike shoes, size 10, navy blue
Not generic recommendation

5. Real-time Process Visualization ⭐

Dashboard shows entire research process:

Agent conversation
Autonomous loop iterations
Citations appearing
Products being configured

Compare side-by-side:

Generic LLM: "Get them a keyboard" (no proof)
Concierge: "Keychron K8 because [3 specific citations from their messages]" (proven)

🔐 Environment Variables

Required

# Agent Authentication
POKE_A_KEY=xxx                          # Gift giver's agent
POKE_B_KEY=yyy                          # Gift recipient's agent

# Supabase
NEXT_PUBLIC_SUPABASE_URL=xxx
NEXT_PUBLIC_SUPABASE_ANON_KEY=xxx
SUPABASE_SERVICE_ROLE_KEY=xxx

# LLM (for agent reasoning)
ANTHROPIC_API_KEY=xxx                   # Claude
# or
OPENAI_API_KEY=xxx                      # GPT

# Browser Automation
BROWSERBASE_API_KEY=xxx
BROWSERBASE_PROJECT_ID=xxx

Optional

# iMessage Query
IMESSAGE_DB_PATH=~/Library/Messages/chat.db
MAX_IMESSAGE_ITERATIONS=10
RELEVANCE_THRESHOLD=0.7

# Poke Integration
POKE_API_KEY=xxx                        # If Poke provides API access

See env.example for complete list.

📚 Tech Stack

Backend

Python 3.12+: Agent logic, MCP servers
fastmcp 2.14.5+: MCP server framework
FastAPI: HTTP transport for MCP
SQLite: iMessage database queries
Supabase: PostgreSQL for application data

Frontend

Next.js 14+: Web dashboard
React 18+: UI components
TypeScript: Type safety
TailwindCSS: Styling
React Query: Real-time polling

Infrastructure

Supabase: Database, auth, real-time
Browserbase: Cloud browser automation
Vercel: Next.js hosting (production)

AI/LLM

Claude (Anthropic) or GPT (OpenAI): Agent reasoning
Agentic Search: Iterative hypothesis validation (no vector embeddings)

🤝 Contributing

This is a TreeHacks 2026 project. For development:

Read the PRD: PRD.md is the ground truth
Understand Agents: See agents.md for architecture
Follow Roadmap: PRD.md Section 10
Update Docs: Keep documentation in sync with implementation

📖 Additional Resources

External Documentation

Internal Documentation

PRD.md - Complete product requirements
agents.md - Agent system architecture
DOCUMENTATION_INDEX.md - Doc navigation
QUICK_REFERENCE.md - Commands cheat sheet

⚠️ Important Notes

1. Ground Truth

Always refer to PRD.md for authoritative specifications. If there's a conflict between documents, PRD.md wins.

2. iMessage Privacy

iMessage database is read-only
Messages are NOT stored in Supabase (only citations for gift context)
Full message access requires user consent (skipped for hackathon demo)

3. Autonomous Agent

iMessage subagent runs autonomously (updates its own hypotheses)
Stop conditions prevent infinite loops
Always returns best candidate even if not fully validated

4. Demo Focus

Prioritize impressive visualizations
Emphasize citation traceability
Show clear differentiation from generic LLMs
Have backup data in case of network issues

🎓 Learning Resources

Understanding Agentic Search

Concierge uses agentic search (not vector embeddings):

Agent tests hypothesis → analyzes results → updates hypothesis → repeat
Example: Test "likes Warriors" → finds "likes Cavaliers" → updates and retests
More transparent and debuggable than embedding similarity

Understanding MCP

Model Context Protocol (MCP) extends AI capabilities via tool servers:

Poke is the base AI assistant
MCP servers provide specialized tools (iMessage search, browser automation)
Agents call MCP tools to perform actions
Similar to function calling but standardized

Understanding Autonomous Loops

Traditional search: Query → Results → Done
Autonomous loop: Query → Analyze → Update Query → Re-query → Repeat until stop condition

📧 Contact

TreeHacks 2026 Team
Built on top of Poke by The Interaction Company

Status: Active Development
Event: TreeHacks 2026
Last Updated: February 14, 2026

🌟 Quick Start Summary

# 1. Clone and install
git clone <repo>
cd treehacks26
uv sync

# 2. Configure
cp env.example .env
# Edit .env with your API keys

# 3. Run MCP server
python mcp/poke-to-poke/server.py

# 4. (Later) Run dashboard
npm run dev

# 5. Watch agents communicate
tail -f conversation.txt

Next Steps: See PRD.md Section 10 for implementation phases.

Analysis

Compare with all teams

View

Metric

Figures cover GitHub contributors during the hackathon window. A co-authored commit counts in full for each author, so per-member totals add up to more than the whole-team figures.

Technology

Found in codeClaimed only

CSSIn code
ExpressIn code
FastAPIIn code
Next.jsIn code
OpenAIIn code
PythonIn code
ReactIn code
Tailwind CSSIn code
TypeScriptIn code

9 of 9 appear in the indexed code.

AI coding agents

CodexConfig

Detected from committed agent config files and commit authorship. Absence of a signal is not proof an agent was unused.

Codebase size

Source size

520 KB

Source files

Counts recognized source files only; vendored directories, binaries and lockfiles are excluded, so this is smaller than the repository on disk.