Handbook - AI Tinkerers - New York City Hackathon
AI Tinkerers - New York City
Sign in

AI Tinkerers - New York City Hackathon

Handbook for the AI Tinkerers - New York City Hackathon hackathon.

Sign in
DevRel Lead - Restate
Build durable AI agent infrastructure. Drive US remote adoption today.
Founding Applied AI Lead - Paxos Health
Build healthcare AI agents in NYC. Apply now.
All Jobs
Final round winners have been announced. View Results

Handbook

ClawHack NYC — Hacker Guide

Theme: Group Agents

1. What This Hackathon Is About

This edition focuses on one specific design space: agents operating inside group chats.

In previous waves, most agents were either:

  • One user interacting with one assistant, or

  • A personal agent with tool access acting on behalf of one user

Today, we are exploring what changes when:

  • Multiple humans share a group thread

  • Multiple agents exist inside that same thread

  • Each agent is scoped to that conversation only

  • Agents have explicit boundaries around visibility and capability

The goal is to experiment with coordination inside a shared conversational context.

The Stack You’ll Use

You’ll be building with:

  • XMTP → secure messaging and identity layer

  • Convos → group chat interface

  • OpenClaw → agent execution runtime

How they fit together:

Layer Role
XMTP Secure communication protocol between humans and agents
Convos UI where the group conversation lives
OpenClaw Runtime where agents reason, use tools, and execute actions

Your project should meaningfully explore group coordination using this stack.

This is not about building a large production system.
It is about exploring interaction patterns under constraint.

2. Schedule and Submission Requirements

Timeline

  • Morning: setup, team formation, stack integration

  • Midday: build

  • Afternoon: testing and demo recording

  • 6:00 PM: submissions close

What You Must Submit

  1. A working prototype
  1. A YouTube demo video that:

  • Shows the group chat in action

  • Demonstrates your agent(s) working

  • Explains what problem you explored

  1. A short written description including:

  • What you built

  • How agents were scoped

  • What coordination pattern you explored

The demo should clearly show how agents operate inside a group thread.

3. What You Can Build

Your project must involve group-based coordination.

Possible directions:

A. Coordinating Agent in a Group

  • Scheduling across multiple participants

  • Coordinating shared logistics

  • Managing shared resources

B. Multiple Agents in One Thread

  • Planner agent delegating to executor agent

  • Research agent + summarizer agent collaborating

  • Agents with different permission scopes

C. Human + Agent Collaboration

  • Agent proposes actions, humans approve

  • Agent mediates or structures discussion

  • Agent only responds when explicitly invoked

D. Scoped / Permissioned Agents

  • Agents restricted to specific message types

  • Agents that cannot access external data

  • Agents that escalate to humans when blocked

The key evaluation question:

How does behavior change when agents operate inside a shared, encrypted group conversation rather than a one-on-one chat?

4. Resources and How They Fit Into Your Stack

Below are the available tools and what role they play.

Convos + XMTP

Communication and identity layer.

  • XMTP provides secure messaging and cryptographic identity.

  • Convos is the group chat interface built on XMTP.

You can:

  • Use Convos Native Assistant to create an agent instantly

    • Write instructions

    • Paste into Convos

    • Deploy into a group chat

You may also connect an existing OpenClaw instance to Convos and use XMTP as the messaging layer.

Use this layer for:

  • Secure message routing

  • Agent identity

  • Group-scoped communication

OpenClaw

Agent runtime.

OpenClaw provides:

  • Tool use

  • Structured execution

  • Controlled file and environment access

Use OpenClaw for:

  • Reasoning

  • Acting on tasks

  • Managing tool boundaries

Your OpenClaw agent should connect to the group via XMTP.

Kilo Code

OpenClaw development tooling.

Kilo can help with:

  • Faster OpenClaw setup

  • Execution scaffolding

  • Tool orchestration patterns

Use Kilo if you want improved development speed around OpenClaw.

Convex

Realtime backend and state layer.

Convex provides:

  • Persistent shared state

  • Realtime updates

  • Structured data storage

Use Convex if:

  • Multiple agents need shared memory

  • You need durable coordination state

  • You want structured backend logic

Zo Computer

Secure compute layer.

Zo provides:

  • Containerized execution

  • Isolated environments

  • Hosted agent runtime

Use Zo if:

  • You need secure execution outside your local machine

  • You want scalable runtime environments

ElevenLabs

Voice interface.

ElevenLabs provides:

  • Text-to-speech

  • Speech-to-text

  • Voice agent capabilities

Use this if:

  • You want voice-enabled agents

  • You want audio summaries or spoken responses

Jelly

AI-native social layer.

Jelly explores:

  • AI-enhanced group interaction

  • Audio-first shared experiences

Useful if your project involves:

  • Social-native AI coordination

  • Audio-driven interaction patterns

5. Suggested Build Paths

Option 1: Fastest Path

  • Use Convos Native Assistant

  • Deploy directly into a group chat

  • Iterate on prompt and behavior

Option 2: Custom Agent

  • Run OpenClaw locally or in a container

  • Connect via XMTP

  • Use Convos as the UI layer

Option 3: Full Stack

  • XMTP for messaging

  • Convos for interface

  • OpenClaw for execution

  • Convex for state

  • Zo for compute

  • ElevenLabs for voice

Choose the level of complexity that fits the time constraint.

6. What a Strong Submission Looks Like

  • Clear scoping of agents

  • Meaningful group coordination

  • Visible interaction inside the thread

  • A clean, understandable demo

Focus on clarity and coordination patterns.

Submissions are due at 6:00 PM with a YouTube demo link.