Agent (Claude Code / Cursor / Codex)
  │
  ▼
contextforge-mcp  ←── single MCP server
  │
  ├── codebase-memory-mcp  (knowledge graph: 99% fewer retrieval tokens)
  ├── headroom             (compression: 60–95% fewer prompt tokens)
  └── Spec Kit             (SDD workflow: spec → plan → tasks → implement)

# Prerequisites
npm install -g codebase-memory-mcp
pip install "headroom-ai[all]"
pip install spec-kit

# Install contextforge-mcp
pip install contextforge-mcp

# Health check
contextforge-mcp doctor

# Configure Claude Code (writes .mcp.json)
contextforge-mcp install --target claude

# Configure Spec Kit extension
contextforge-mcp install --target speckit

# Both at once
contextforge-mcp install --target all

# 1. Index the codebase (once per session)
cbm_index_repository(repo_path=".")

# 2. Query the graph (instead of grep/read)
cbm_search_graph(name_pattern=".*Payment.*")
cbm_trace_path(function_name="processPayment")
cbm_get_architecture()

# 3. Check token savings
cf_stats()

/speckit.constitution
/speckit.specify
/speckit.cf-analyze      ← analyze codebase before planning
/speckit.plan
/speckit.tasks
/speckit.cf-index        ← index before implementing
/speckit.cf-implement    ← graph-aware implementation
/speckit.cf-stats        ← token savings report