nexlab

services

Your API, callable from Claude, Codex and any other LLM

Model Context Protocol servers, built and run by someone who sells 23 of them

An MCP server is the layer that lets an AI assistant use your product instead of describing it — from Claude, Codex, ChatGPT, Cursor, or any client that speaks the protocol. Done well, it gives an agent exactly the access you decide and nothing more: reads and writes as separate tools with separate scopes, the scope fixed when the token is issued, quotas enforced at the edge. Your customers get your product inside the tools they already use; you keep the boundary. We design it that way, deploy it where you want it, and run it if you would rather not.

▸ info@nexlab.net ▸ Book a call See the servers live

sized to the job

🧩

Simple server

Up to five tools over one public API; key or no auth; read-only. Typed schemas, validation at the boundary, a test suite and a README, deployed where you want it. Two to three days.

🛠️

Standard server

Up to a dozen tools over one or two sources; OAuth or per-user keys; writes with proper scoping; rate limits, health checks, errors that degrade cleanly. About a week, with 30 days of fixes.

🏗️

Platform

Several sources behind one coherent tool surface; multi-tenant auth, secrets handling and observability; read-only, confirm-required and forbidden actions; optional marketplace listing. Scoped after an audit, with 60 days of fixes.

🔍

Audit from €250

The one number on this page. Half a day for a simple server, longer for a platform: tool surface, schemas, auth and scoping, and what happens on malformed calls and expired credentials — written findings and a fix plan. Deducted from the build if you go ahead within 30 days. Everything else is quoted within a day after a short call.

Why me, checkably

Anyone can say they build MCP servers. These are things you can verify before writing to me.

mcp.nexlab.net — 23 MCP servers with 200 tools, sold as a hosted service behind one endpoint and one key. OAuth 2.1 with dynamic client registration and PKCE, metered billing in EUR with card, PayPal and crypto rails, and typed errors an agent can act on. Every answer names its sources; a fleet-wide conformance check that calls every tool with synthesised arguments currently reports 0 violations across 200 calls.

Listed where agents look. The servers are in the official MCP Registry under a DNS-verified namespace, on the MCPize marketplace and on Apify. The same code runs on all three, deployed from one source with one script.

Run against real agents. These servers have been driven by Claude, Cursor and ChatGPT-style clients in production, which is where you learn how models actually fail: malformed tool calls, truncated JSON, hallucinated tool names, confident action on stale state. Retries with backoff, bounded caches, hard deadlines on every upstream call and hints on every error exist because each of those happened.

what they cover

planet-watch — 41 toolssky-watchpeer-research-paperssacred-textscourt-recordseu-policy-watchus-congress-watchcorporate-money-trailtrade-tariffsvulnerability-intelshodan-searchflight-watchmaritime-aiseurope-rail-watchlaunch-watchgrid-carbonwikipedia-retrievalyoutube-media+5 more

and underneath it

  • clawphone — a job-queue MCP server coordinating several agents: concurrency, claim races, cleanup after dead workers.
  • hermes-node-gateway — the runtime and routing between an agent and its device nodes, including a Chrome node driving a real browser and Android nodes in Kotlin.
  • CoderAI / AISBF — local model orchestration behind an OpenAI-compatible API across NVIDIA, AMD and Intel, with multi-provider routing and failover.
  • 25+ years of production Linux. Co-founder of Devuan, author of Amprolla.

Who this is for

  • "Our customers use Claude. They should be able to use our product from inside it." SaaS companies whose API exists but has never been exposed to agents safely.
  • "We have data nobody can query." Data providers, registries and research organisations with APIs or databases and no agent surface.
  • "Our internal tools are a mess of scripts." Ops teams who want their runbooks callable by an assistant, with proper scoping.
  • "We built an MCP server and it keeps breaking." The rescue job — start with the audit.
  • "People have already built MCP servers on our API." Good sign — there is demand. We make the official one, with your scopes, your name and your support path, and the audit starts with an inventory of what already exists so nothing your users rely on is lost.

How it goes

  1. Audit first whenever something already exists. It is cheap enough to approve without a procurement cycle, and the report is the build's specification.
  2. Build against your real API with your real auth model, deployed where you want it — self-hosted, a marketplace, or behind a gateway like the one at mcp.nexlab.net.
  3. Run. Every deliverable ships with tests, a runbook and a fixes window; ongoing maintenance, sized to what is being maintained, is there when you would rather not own the upkeep.

Questions people ask

Which clients? Anything that speaks MCP — Claude Desktop and Claude Code, OpenAI Codex and ChatGPT, Cursor, Windsurf, VS Code, and agent frameworks. Streamable HTTP with stateless sessions, so it scales without session affinity; stdio for local use.

Self-hosted or marketplace? Either. I have shipped both, and I will tell you when a marketplace's cold start or revenue share makes owning the edge the better deal.

How much? The audit starts at €250 and comes off the build. Builds and ongoing maintenance are priced to the size of the job — a five-tool server and a multi-tenant platform are not the same work — and you have a fixed quote within a day of a short call.

Where are you? Italy and South Africa; remote, worldwide. Invoices in EUR or USD.

$ ./expose-your-api-to-agents

Tell me which API, which clients your customers use, and what an agent must never be allowed to do. I reply with a scope and a price, usually the same day.

▸ info@nexlab.net ▸ Book a call Read about mcp.nexlab.net