id: RD-0028 title: "Context7 MCP — Live Library Documentation Injection" requestor: CEO (Oscar) assignee: R&D Analyst (Riley) status: open created: 2026-05-12 completed: impacts: [intranet, frontend-design, web-designer] type: rd-evaluation decision: decision_date: decision_by: outcome: outcome_date: references: [RD-0003]

RD-0028 — Context7 MCP Evaluation

Evaluate the Context7 MCP server as a tool for injecting up-to-date library documentation into Claude Code sessions. Surfaced as hack #32 ("game-changer") in RD-0003.

Why now

Intranet runs on Astro (drift Astro 4 → 5 documented in project_intranet).
frontend-design skill regularly suggests outdated React/Tailwind APIs.
agent-ea v2 regen pending (project_agent-ea-v2) — will touch lib-heavy code.

Hypothesis to test

Context7 injection improves V1 quality on lib-touching tasks enough to justify its context-token cost in the targeted talents — not as a factory-wide MCP, but as a conditional activation on specific digital talents.

Sources

Original signal: RD-0003 hack 32 — https://www.youtube.com/watch?v=jqoFP9QapXI
Synthesis context: references/videos/RD-0003/out/synthesis-strategique_rd-0003.md §1 Tier A
Install ref: Context7 (upstash) — confirm canonical install command via Context7 docs

Evaluation criteria (R&D pipeline TFD-0012)

Riley to score on the standard 5 dimensions; pay attention to:

Token cost overhead per session (Context7 charges into context window — hack 24 warning)
V1 quality delta on Astro intranet (concrete test case)
Maintenance burden vs adding/removing per-talent

Anchoring

project_intranet — Astro version drift
project_framework-library — Macroscope/LeanIX frameworks
project_rd-pipeline — adds to active R&D queue