- Client: AI Agent Product, an AI × Web3 tooling company
- Category: AI Agents / DeFi Tooling
- Engagement: Dedicated team, AI + Web3 engineering
- Duration: 2 months, concept to launch
- Key result: 40K+ users, 30+ DApps integrated via chat
AI Agent Product: a chat interface for crypto research and on-chain actions, built to catch the AI narrative early.
The client wanted to build an AI agent that made Web3 easier for everyday users, but Web3 today asks a lot of its users: switching between block explorers, DEX interfaces, and news sources just to answer a simple question or make a trade. The client needed a product that collapsed all of that into one place, and needed it built fast enough to launch while the AI agent narrative was still gaining momentum.
The client had explored a couple of chatbot-wrapper approaches internally and hit a wall the moment they tried to connect the chat to real on-chain actions. They chose us for our experience combining LLM tool-calling with Web3 infrastructure - specifically our approach to MCP-based integrations - and for a 2-month delivery plan that matched how quickly the AI agent narrative was moving.
Two parallel workstreams from week one, merged in the final three weeks.
1 Product Manager, 1 Business Analyst, 1 Tech Lead, 2 AI/LLM engineers, 2 Web3 backend engineers, 1 frontend engineer, 1 DevOps Engineer, 1 QA engineer.
Conversational research features and MCP/on-chain action integration ran in parallel from week one, merged in the final three weeks.
Weekly product demos so the client could test real conversational flows early, rather than reviewing a spec.
The product centers on a chat interface backed by an LLM with tool-calling, where every capability - price lookups, wallet insights, news retrieval, and on-chain actions - is exposed through MCP (Model Context Protocol) servers rather than hardcoded into the chat logic. New capabilities could be added as new MCP tools without retraining or restructuring the core agent.
LLM-backed chat with tool-calling for coin search, portfolio insights, and real-time news aggregation.
Purpose-built MCP servers wrapping DEX aggregator, bridge protocols, and liquidity contracts. Every on-chain action requires an explicit transaction preview and wallet-side confirmation before signing.
WalletConnect/viem integration for signing, with transaction simulation before every confirmation prompt. React chat interface with streaming responses and inline transaction cards.
Tech stack: LLM API with tool-calling · Model Context Protocol (MCP) · Node.js/TypeScript · DEX aggregator API · WalletConnect/viem · React
We built the MCP integration as its own well-defined layer between the chat model and Web3 actions: the LLM can only call a fixed set of structured tools with typed parameters, never freeform transaction data - closing off an entire class of prompt-injection and hallucinated-transaction risk.
Every action runs through transaction simulation and an explicit human confirmation step before signing. That kept the product simple for users while keeping every on-chain action safe and intentional.
The safety architecture didn’t slow the build - it was designed in from week one. Structured tool-calling and simulation checkpoints were part of the architecture, not a post-launch patch.
Chat UI, research/insights tool-calling, MCP server architecture.
Swap/bridge/liquidity MCP integrations, transaction safety layer, testing, launch.
The client launched a working AI agent that lets users research and act on crypto - swap, bridge, add liquidity - from a single chat interface, in time to capture the AI agent narrative early. Within the first quarter, the product had integrated with 30+ DApps and crossed 40,000 users, handling 25,000+ agent-executed actions with zero incidents of unintended transaction execution.
Read the full category overview, market data and outlook.