Your product is invisible to AI agents.
Claude has a vetted connectors directory. ChatGPT launched with 1,400+ connectable apps. When a customer asks their AI assistant to use your product and nothing happens, they end up on whichever competitor built the integration. I build MCP servers with the auth, logging, and submission work it takes to pass directory review and stay reliable.
Why me
I built the side of the stack agents talk to.
I spent two years at ZeroClick, a company whose pitch is “the OS for selling to agents,” writing the agent-facing side of the stack: the APIs and MCP servers that AI agents consume, the developer documentation that integrators build against, and Activation Ads, an agent-native ad format I originated. Before that I spent five years at PayPal building Honey at consumer scale. I learned the protocol shipping to agents in production, and I use agent tooling every day, so I know a badly designed tool surface when an agent trips over one.
Start here
Know what agents need from you.
- Review of your API for agent consumption: what agents need that it doesn’t expose
- Designed MCP tool surface: the tools, their schemas, and the descriptions that make agents use them correctly
- Auth architecture recommendation (OAuth flows, scopes, what an agent may do on a user’s behalf)
- Claude and ChatGPT directory requirements gap list
- Build-vs-buy recommendation and a fixed-bid quote for the implementation
What builds cost
The audit produces a fixed bid. For reference, agencies quote $25k–$50k for SMB implementations and $60k–$120k for production builds. I typically come in under both because you hire one engineer.
| MCP server for an existing SaaSTool-surface design, OAuth with scoped permissions, remote hosting, logging, tests, and directory submission for Claude and ChatGPT. | 4–7 weeks | $18k–$35k |
| Prototype productionYour internal MCP demo hardened into something that survives real auth, rate limits, and multi-tenancy, without a rewrite where possible. | 2–3 weeks | $8k–$15k |
| MCP + usage meteringFor API-first companies that want agents as paying customers: the MCP surface plus metering, quotas, and billing integration. | 6–10 weeks | $30k–$60k |
How working with me goes
- Audit
Fixed price, one week. Tool-surface design, auth plan, and directory gap list. It is useful even if you build in-house.
- The build
Fixed bid. I implement the MCP server in your repos: remote-hosted, with OAuth, rate limiting, logging, and tests.
- Listing & handoff
I drive the Claude and ChatGPT directory submissions, document everything, and train your team to extend the tool surface themselves.
Guides
Read before you hire anyone
Written from running MCP servers in production, not from the spec. If a guide solves your problem outright, you don’t need me.
- Your MCP server is stateful. Your load balancer doesn't know that.Sessions, sticky routing, proxy timeouts, and what breaks at two replicas.
- MCP went stateless. Here's how to migrate a session-era server.The 2026-07-28 spec removes sessions. The handle pattern, elicitation changes, and the order to migrate in.
- How to pause an MCP tool call, ask the user, and resumeMCP elicitation end to end: suspend and resume, timeouts that ignore human thinking time, and the race everyone hits.
- Why models ignore your MCP tools, and the design that fixes itServer instructions, filtering server-side, error results over exceptions, and tool output as prompt surface.
- You have to run untrusted JavaScript. A VM sandbox isn't enough.A VM context isolates scope, not capability: SSRF and rebinding defenses, log exfiltration, and the execution contract.
Questions clients ask
What does a production MCP server cost?
Published market rates for production-grade MCP implementations run $25,000–$120,000 through agencies. My builds typically land under agency pricing because there’s no bench and no project-manager layer. The audit gives you a fixed number for your scope.
MCP, ChatGPT apps, or both?
They share most of their design thinking, and MCP increasingly works across both ecosystems. The audit answers this for your product, including whether one MCP server covers both, which it often does.
How do we keep an agent from doing something destructive with a user’s account?
Scoped OAuth, capability-level permissions, human confirmation for irreversible actions, and audit logging. This is the part prototype integrations skip and security teams block. I design it in from the first day.
Can’t we just point agents at our existing REST API?
You can. Agents will fumble it. Tool design for agents is different from API design for developers: fewer tools with clearer purposes, descriptions the model can follow, and responses shaped for reasoning instead of parsing.
Who owns the code? Will you sign an NDA?
Everything is work-for-hire, in your repos, on your infrastructure. You own all of it. NDAs are fine.
Other things I’m hired for
Safari & iOS extension porting
I port Chrome extensions to Safari, macOS, and iOS, the way I did at Honey and Pie.
Learn moreAI-Agent Development Enablement
I get engineering teams shipping production software with AI coding agents.
Learn moreAI-Built POC to Production
I take AI-built prototypes to production: tests, auth, data, deployment.
Learn moreYour customers are already asking their assistants for this.
Send me a link to your API docs and a sentence about what your customers want their AI assistants to do. I'll tell you within a day what an MCP integration would look like.