What the work involves

MCP is the connective tissue between a language model and the tools, files, and APIs it acts on — and most of the failure modes in agentic systems live in that seam. On a typical engagement you are writing MCP servers (resources, tools, prompts) in TypeScript or Python, wiring them to a host like Claude Desktop, Cursor, or a custom agent runtime, and then producing the evaluation material that tells a lab whether the model used them correctly. That means authoring tool schemas that are unambiguous to a model, constructing tasks where the correct behaviour requires two or three chained calls, and writing rubrics that separate "reached the right answer" from "reached it by a legitimate sequence of calls."

A large share of the day is judgment rather than code. You will be reading transcripts where a model invented an argument the schema never declared, silently retried after a 500, called a destructive tool without confirmation, or hallucinated a resource URI. Your job is to label what went wrong, attribute it to the model versus the server versus the prompt, and often patch the server so the failure is no longer possible. Some projects lean toward server implementation and hardening; others are almost entirely trajectory review and rubric writing.

What the platform screens for

micro1's process opens with an AI interviewer that asks about your background and then pushes on the specifics — expect follow-ups on transport choices, why stdio versus streamable HTTP, how you handle authentication and secrets in a server, and what you did when a tool description caused a model to misfire. Generic agent-framework experience is not the same thing as MCP experience, and the screen is built to tell them apart. Live or recorded coding is common, and claims on your profile are re-probed later in the process.

Logistics

  • Remote, contractor engagements; most work is asynchronous with occasional sync calls in US-overlapping hours.
  • Part-time commitments of roughly 10–20 hours a week are typical, with some projects scaling to full-time.
  • Observed rates for this role run $60–120/hr, varying by depth of MCP work and project; treat the band as reported, not promised.
  • Own machine, a GitHub account with reviewable history, and comfort working against fast-moving, sometimes undocumented spec revisions.