— Practice / MCP
MCP server development.
The Model Context Protocol is how AI agents call your software as tools: search your data, run an action, finish a task. We design and build production MCP servers on Cloudflare Workers, deployed globally, with the input validation, auth, and rate limits a demo skips.
— Why us
We run the thing we sell.
This site runs a public, stateless MCP server on the current SDK v2 stack. It negotiates the current protocol with modern clients, preserves a stateless compatibility lane for legacy clients, validates Host and Origin, rate-limits by request, and exposes six working tools. You can inspect it rather than taking a capability claim on faith.
- Protocol
- MCP 2026-07-28
- Transport
- Stateless HTTP
- Public tools
- 6
- Compatibility
- Modern + legacy
Connect a remote MCP client to https://setkernel.com/mcp. No auth needed. Discovery metadata is public at https://setkernel.com/.well-known/mcp/server-card.json.
— What we build
From one tool to a full agent surface.
Tool servers
Expose a database, an internal API, a search index, or a booking flow as MCP tools an agent can call safely, with input validation and rate limits.
Streamable HTTP transport
Modern remote MCP over HTTP (the transport Claude and Cursor use), deployed on Cloudflare Workers so it scales to zero and runs close to every user.
Auth and access control
Bearer tokens, OAuth, or fully public, scoped to what the tools actually need. Production also needs an audit trail and revocation, so those come with it.
Grounded retrieval
Pair the server with a RAG pipeline so agents answer from your own content, with citations. We build both halves.
Evals and guardrails
Tests on tool behaviour, structured logging, and enough observability to see what agents actually did. This is what keeps an agent honest in production.
Server card and discovery
A published server card and the well-known discovery surfaces so agents and directories can find and describe your server without a human in the loop.
— How it works
The engagement.
Same five-step method as every SetKernel build (Brief, Architect, Sprint, Ship, Operate), each with a written artefact you review. We start from a short written brief: which systems the agent should reach, who should be allowed to call them, and what done looks like. You get a scoped price and a fit / no-fit answer within one business day.
— Where we work
Atlantic Canada, and worldwide.
We are a complete technology partner in Halifax, Nova Scotia, Canada, and we work remotely with teams well beyond the region. MCP runs over HTTP, so where your team sits does not change the build.
— Questions
Before you write.
What is an MCP server, in one sentence?
A Model Context Protocol server exposes your software as a set of tools that an AI agent (Claude, Cursor, ChatGPT, or any MCP client) can call to look something up or take an action. The protocol is standard, so you build it once and every client works.
Why build it on Cloudflare Workers?
Workers scale to zero, run in every region by default, and have no cold-start tax, so an MCP server costs almost nothing at rest and stays fast under load. It is also the platform we know best, and we build for it natively instead of porting a container.
Do you only build the server, or the agent too?
Either. Some clients have an agent and need a server to give it tools. Others want the whole system: server, retrieval, evals, and the agent that uses them. We scope to what you actually need.
Can we see your work first?
Yes. Install the MCP server that powers this site, read the open-source cf-mcp-template, and its companion essay on running MCP in production. Working code says more than a pitch.
How do we start?
Send a short written brief: the systems the agent should reach, who may call them, and the deadline. We reply in writing within one business day with fit / no-fit and, if fit, a scope and price. No discovery call before the brief.
— Engage
Have something you want an agent to be able to do?
Tell us in two paragraphs: the systems, the callers, the deadline. We reply in writing within one business day.