What an MCP Developer does
An MCP developer is a senior backend or platform engineer who implements the Model Context Protocol: the open standard that lets an AI client such as Claude or Cursor call your systems through a declared set of tools instead of a pile of bespoke glue. The deliverable is a server. It declares which tools exist and what arguments they take, authenticates the human behind the request, scopes every call to what that person is allowed to see, rate-limits it and writes it to an audit log. Teams usually discover the protocol takes days to learn and that the real work is auth, scoping and auditability. That is why this is a security-minded backend hire, not an AI hire.
Monthly rate
$6,000–$9,000/mo
All-in: contract, benefits, equipment, IP
Experience
10+ years engineering
API and auth depth
Location
Latin America
Argentina · Colombia · Mexico
Timezone
Full US overlap
Fluent English, onboarded in one week
Core stack
AI tools, daily
Verticals seen
What an MCP Developer owns - and what they don't
What they own
- Build MCP servers that expose your APIs, databases and internal services as typed tools with explicit contracts
- Implement auth end to end: OAuth flows, per-user credential isolation, least-privilege scopes, refresh and revocation
- Add rate limits, timeouts and idempotency keys so an agent stuck in a loop cannot take a production system down
- Instrument audit logging: which user asked, which tool ran, with what arguments, against which record
- Package and distribute the server: remote or local, versioned and documented, for your team or your customers to connect
What they don't - and who does instead
- Build product features on top of models (RAG, evals, chat flows). That's an AI Engineer
- Design the multi-agent system that consumes the tools. That's an AI Agent Architect
- Attack the finished surface adversarially. That's AI Security & Red Teaming
- Set up your team's day-to-day coding workflow. That's a Claude Code Developer
Who hires an MCP Developer, and for what
Platform teams opening systems of record to agents. The CRM, the warehouse, the internal billing API. The blocker is never the protocol. It is proving that a given agent call can only ever reach one user's data.
SaaS vendors publishing an MCP server for their customers. Buyers now ask whether you have one during evaluation. That makes it a supported, versioned, documented surface held to the same bar as your public API, not a weekend project.
Enterprises with a working assistant and no integration story. It answers questions and cannot act, because security will not grant credentials until scoping and audit logging exist. That gap is this role's entire job.
- 01
First internal MCP server. One system, a handful of read-only tools, OAuth and audit logging from the first commit. Typically three to five weeks with one engineer.
- 02
Going from read to write. Tools that create records, update state or move money: idempotency keys, confirmation steps on irreversible actions and a rollback path. Six to ten weeks depending on how many systems are in scope.
- 03
Customer-facing MCP server for your product. Remote server, OAuth against your identity provider, per-tenant scoping, docs and a version policy so client connections do not break on release. Planned like a public API launch.
- 04
Hardening a server you already shipped. The hackathon version works and holds one shared admin token. Re-scoping credentials per user, adding audit logging and closing the injection paths that tool access opens runs two to three weeks when the tool surface stays as it is, longer if the server has to be redesigned.
Work our engineers at this role have shipped
Greelow has been building backend and integration work since 2017 for clients including OneReach.ai, Banorte and Wallbox, rated 5.0 on Clutch and ISO 9001 certified. The MCP work is more recent and its case study runs unnamed at this stage: an agentic operating system for a private-markets firm, three quarters in production with no credential-scope incidents. The engineers are the same ones.
- Catalog of MCP servers for a Claude- and Cursor-native enterprise buyer: one typed tool surface per internal system, versioned and documented so teams connect without bespoke glue
- Per-user OAuth credential isolation on a multi-tenant agentic platform for a private-markets firm, so every tool call carries the identity of the human who triggered it
- Tool permission boundaries and cross-tenant leakage tests closed on an agentic platform before its tools were opened beyond the internal team
- Webhook and real-time notification layer for a B2B SaaS wired to dozens of third parties: the same auth, rate-limit and idempotency work a write-capable MCP server needs
Custom MCP server development vs off-the-shelf connectors
Connectors cover commodity SaaS somebody else maintains. A custom server covers the systems only you have.
| Custom MCP server | Off-the-shelf connector | |
|---|---|---|
| Systems it reaches | Your internal APIs, proprietary data, your own product | Commodity SaaS with a published integration |
| Permissions | Per-user OAuth, scoped to what that person may see | Usually one shared service account |
| Audit trail | User, tool, arguments and target record, logged by you | Whatever the vendor chose to expose |
| Time to first working tool | Three to five weeks, one system, read-only | Same day |
| Who maintains it | Your team, versioned like a public API | The vendor, on their release schedule |
Do you actually need an MCP Developer?
You do, if:
- Your assistant or agent can read but cannot act, because nobody will hand it credentials
- Every new integration is bespoke glue plus another secret in a config file
- Customers or internal teams are asking for an MCP server and no one owns it
- You have a server in production running on a shared service account with no audit trail
You probably don't, if:
- You need LLM features in the product itself (RAG, evals, chat). Hire an AI Engineer
- The open question is how agents coordinate, not how they reach your systems. An AI Agent Architect designs that layer
- Off-the-shelf connectors already cover your tools and nobody needs per-user scoping yet. You don't need a custom server
Not sure which role fits? Tell us the problem instead of the title - we'll tell you what we'd actually staff, even if it's not this. If it is this: discovery call today, matched profiles in 48 hours, onboarded in a week.
Need more than one role, or a whole team around this hire? See how nearshore staff augmentation works. Building this across your whole product, not just one hire? See our AI engineering service.
Hire a Senior MCP Developer