Skip to main content

MCP Context7 Integration

On-demand library documentation via Context7 MCP

Goal​

Integrate Context7 MCP to fetch, cache, and manage library documentation for development tasks.

Skill: /dev-docs-fetch​

  • Name: dev-docs-fetch (role=dev, domain=docs, action=fetch)
  • Directory: skills/dev-docs-fetch/
  • MCP Required: Context7 (mcp__context7__resolve-library-id, mcp__context7__get-library-docs)
  • Input: [library-names...] or auto-detect from $JAAN_CONTEXT_DIR/tech.md
  • Output: Cached docs in $JAAN_OUTPUTS_DIR/dev/docs/context7/

Capabilities​

FeatureDescription
Resolve library IDsmcp__context7__resolve-library-id
Fetch documentationmcp__context7__get-library-docs (mode: code/info)
Smart caching7-day TTL, Bash-based freshness check
Auto-detectionRead $JAAN_CONTEXT_DIR/tech.md for stack-aware library resolution
Fallback handlingUse stale cache on API errors, retry with backoff
Topic extractionExtract specific topic from task description (e.g., "middleware", "streaming")

Subtasks​

  • Add Context7 MCP server to Claude settings (.mcp.json)
  • Create /dev-docs-fetch skill (skills/dev-docs-fetch/SKILL.md)
    • YAML frontmatter with allowed-tools: Bash, Context7 MCP tools, Write, Read
    • argument-hint: [library-names...]
    • Pre-execution protocol integration
    • Two-phase workflow with HARD STOP
  • Implement tech-agnostic library detection
    • Read $JAAN_CONTEXT_DIR/tech.md for current stack (no hardcoded library tiers)
    • Parse arguments when provided directly
    • Auto-detect from conversation context as fallback
  • Implement cache freshness check (Bash-based)
    • Cache directory: $JAAN_OUTPUTS_DIR/dev/docs/context7/
    • 7-day TTL using file modification time
    • Cross-platform (macOS stat -f %m / Linux stat -c %Y)
  • Add graceful error handling
    • Library not found → suggest alternatives, offer retry/skip
    • API failure → fallback to stale cache
    • Network timeout → retry up to 3 times with backoff
    • Invalid response → skip and continue
  • YAML frontmatter for cached files
    • Fields: title, library_id, type, created, updated, context7_mode, topic, tags, source, cache_ttl
    • Preserve created date on re-fetch
  • Learning integration ($JAAN_LEARN_DIR/dev-docs-fetch.learn.md)

Skill Workflow​

Phase 0: Parse arguments + auto-detect from tech.md
Phase 1: Cache freshness check (fresh/stale/missing)
   ↓
HARD STOP — Confirm libraries to fetch
   ↓
Phase 2: Resolve library IDs via Context7 MCP
Phase 3: Fetch documentation (code/info mode)
Phase 4: Store with YAML frontmatter + summary report

v3.0.0 Adaptation Notes​

AspectDesign Decision
Tech-agnosticNo hardcoded library tiers — read $JAAN_CONTEXT_DIR/tech.md for stack detection instead
Env varsAll paths use $JAAN_* variables ($JAAN_OUTPUTS_DIR, $JAAN_CONTEXT_DIR, $JAAN_LEARN_DIR)
Cache path$JAAN_OUTPUTS_DIR/dev/docs/context7/ (follows output structure convention)
Learning$JAAN_LEARN_DIR/dev-docs-fetch.learn.md with merge strategy
Multi-runtimeSkill definition is runtime-agnostic (canonical in skills/)

File Structure​

$JAAN_OUTPUTS_DIR/dev/docs/context7/
├── fastapi.md
├── react.md
├── sqlalchemy.md
└── ...

Each file includes YAML frontmatter:

---
title: FastAPI Documentation
library_id: /tiangolo/fastapi
type: context7-reference
created: 2026-02-22
updated: 2026-02-22
context7_mode: code
topic: null
tags: [context7, fastapi, technical-reference]
source: Context7 MCP
cache_ttl: 7 days
---

Integration Points​

  • Standalone: /dev-docs-fetch fastapi openai
  • Auto-integrated: Can be called from other dev skills in Phase 1
  • Silent mode: Cache hits don't produce output

Token Budget​

Target: <10,000 tokens per execution

  • Check cache with Bash (no file reads for fresh cache)
  • Fetch only stale/missing libraries
  • Max 3-5 libraries per run
  • Concise reporting (no verbose logs)

References​