AI Agents wired into
the systems you already run
Custom MCP servers, RAG knowledge assistants, and production hardening for Hong Kong businesses. Connect Claude, Cursor, or any MCP-capable AI to your POS, CRM, database, or internal API - with the guardrails that make it safe to leave running. Briefed and built by the same person.
We don't just build these - we run one
HiMarket, our own trading-card marketplace, exposes a public MCP server that lets any AI agent trade as its account owner: search the catalogue, read live market prices, and manage sell listings and buy offers - with real store credit behind every write.
- Full marketplace toolset: catalogue search, live market data, and the account’s own listings
- Scoped tokens (read / sell / buy) with a propose mode - third-party agents draft, the owner approves
- OAuth discovery and personal access tokens
- Budget caps and per-token rate limits guarding real store credit
- Listed in the official MCP registry as hk.himarket/mcp
Every guardrail we quote for your system is one we already operate with our own money on the line.
Read the live MCP documentationWhat We Build
The pieces of agent work we have production experience with - and how they fit the systems Hong Kong businesses actually run.
Custom MCP servers
Give Claude, Cursor, or any MCP-capable client working hands on your systems - stock lookups, order status, price updates, bookings. Your data, your permissions, one standard connector instead of a one-off integration per AI tool.
RAG knowledge assistants
Assistants that answer from your own documents - SOPs, product sheets, manuals, filings - in Cantonese and English, with citations so staff can verify every answer instead of trusting a black box.
Guardrails & production hardening
Scoped tokens, propose-then-approve write flows, budget caps, rate limits, audit logs. Already built 70% of an agent with AI yourself? We do the last 30% that makes it safe to leave running.
Legacy & no-API systems
SQL POS databases, ODBC, on-premise vendor software, undocumented APIs. The systems AI tools cannot introspect on their own are most of the real work - and most of what we get asked to connect.
Assistants where your customers are
The same agent work delivered on the channel Hong Kong actually uses. Our own WhatsApp number is answered by one - see the WhatsApp Chatbots line.
Document processing
Extraction and classification that ends in structured data inside your systems - invoices, forms, filings - not a chat window someone has to copy-paste from.
Why our agent work is different
Proven on our own systems
Our public MCP server, our WhatsApp assistant, and our internal research tools run in production first. You get patterns that survived real users, not a slide about them.
Direct to the developer
No project manager telephone game. The person who maps your workflow is the person who writes the connector, ships it, and answers when something goes sideways.
Fixed-fee, scoped
A written scope and price before the build starts, and clean source you can hand to anyone afterwards. No proprietary platform, no lock-in to us.
How a typical engagement looks
A short call to map one painful workflow: which systems hold the data, who needs the answers, what the AI should be allowed to touch. You get a written scope and a fixed-fee proposal within days.
A real MCP server or assistant running against your actual system - not a mockup. You point Claude or Cursor at it and see your own data answering.
Auth, scoped permissions, approval flows, rate limits, logging, deployment your team can run. Source and documentation handed over at the end.
Frequently Asked Questions
Straight answers for owners and technical leads considering agent integration, MCP servers, or a RAG assistant.
What is an MCP server?+
MCP (Model Context Protocol) is the open standard that lets AI tools such as Claude, Cursor, and ChatGPT connect to outside systems. A custom MCP server is the connector that exposes your database, POS, CRM, or internal API to those tools - under permissions you control. Build it once and every MCP-capable AI client can use it.
Can an AI agent write to live systems safely?+
Yes, if the guardrails are designed in rather than bolted on. On our own marketplace MCP server, writes are gated by scoped tokens, a propose mode where the agent drafts and a human approves, budget caps that abort a run before it overspends, and per-token rate limits. We build the same discipline into client systems.
Our system has no modern API - can you still connect it?+
Usually, yes. SQL and ODBC databases, on-premise vendor software, and undocumented APIs are the systems AI tools cannot reach on their own - and they are the majority of what we are asked to connect. If it stores data, an agent-friendly interface can normally be put in front of it.
Which AI tools do you support?+
Anything MCP-capable: Claude, Cursor, ChatGPT-class clients, and self-hosted models. Knowledge assistants can be delivered on WhatsApp, as a website widget, or as an internal tool - the connector and the knowledge base are reusable across all of them.
Do we own the result?+
Yes. You receive the source, deployment notes, and documentation. No proprietary platform, no lock-in - keep us on retainer or hand it to your own team.
Agent work is one part of our Web Technology line — often paired with:
Looking for our Crestron smart home work instead? See Smart Home →