linear-mcp

A production-grade, open-source MCP server connecting Claude to Linear, plus a reference agent that completes a real multi-step task across it.

Target role this answers: Forward-Deployed Engineer — this is the literal artifact Anthropic's FDE JD asks for: "deliver technical artifacts for customers like MCP servers."

Pillar: Integration + hardening — a public, verifiable artifact (GitHub + npm) with engineering rigor most candidates never show.

🏗️ Build status: scaffolding. Project 2 of 3. Also doubles as the escalation backend for support-copilot.

---

What it does

A TypeScript MCP server on @modelcontextprotocol/sdk exposing 6–10 typed tools over the Linear API (search / get / create / update / comment, list teams / projects), wrapped in a real resilience layer.

Architecture (planned)

• Tools — zod-typed input + output schemas for each tool.
• Resilience layer — OAuth/token handling, rate-limit handling, retries with exponential backoff, pagination, idempotency, structured recoverable errors.
• Reference agent — Claude with tool use completing a real cross-system task end to end.
• Quality — vitest suite + tool-use evals (right tool, right args, across N scenarios; reuses the EvalOps harness) + GitHub Actions CI.
• Distribution — published to npm + listed in an MCP registry.

Tech stack

TypeScript · @modelcontextprotocol/sdk · zod · Vitest · Claude API

Install (planned)

npx linear-mcp        # one-line install, once published

Definition of done

Public repo + npm package + registry listing, passing CI, a tool-use eval suite, and a Loom of Claude completing a live multi-step task.

Interview-ready checklist

• [ ] Public GitHub repo + README with architecture diagram
• [ ] npm package + MCP registry listing
• [ ] Green CI + tool-use eval suite
• [ ] 60–90s Loom of a live multi-step task

License

MIT — see LICENSE.

---

Part of Road to First AI Role. Reuses the eval harness from EvalOps; is the escalation backend for support-copilot.