MCP 服务器

Pharaoh

AI reads your codebase like a book - one page at a time. Pharaoh gives it the index and a map of how every chapter connects.

README

Pharaoh

Codebase intelligence for AI agents. Your AI understands your architecture before it writes a single line.

pharaoh.so

Quick Start (60 seconds)

Claude Code

claude mcp add pharaoh --transport sse https://mcp.pharaoh.so/sse

Cursor

Add to .cursor/mcp.json:

{
  "mcpServers": {
    "pharaoh": {
      "url": "https://mcp.pharaoh.so/sse"
    }
  }
}

Generic MCP Client

{
  "mcpServers": {
    "pharaoh": {
      "transport": "sse",
      "url": "https://mcp.pharaoh.so/sse"
    }
  }
}

Setup

Add the MCP URL to your client (see above)
Authorize via OAuth — you'll be prompted to install the Pharaoh GitHub App on your org
Repos are mapped automatically. Start querying.

What Pharaoh Does

Pharaoh parses your repositories using tree-sitter into a Neo4j knowledge graph — functions, modules, imports, exports, call chains, endpoints, complexity scores, all mapped as nodes and edges. AI agents query the graph via MCP for structured architectural context instead of reading files one at a time. Deterministic analysis — no LLM in the pipeline, zero hallucination risk.

Tools

Tool	What it does	Tier
`get_codebase_map`	Full structural overview — modules, dependency graph, entry points, hot files	Free
`get_module_context`	Complete module profile — functions, exports, internal deps, complexity scores	Free
`search_functions`	Find existing functions before writing new ones — name, signature, location, callers	Free
`get_blast_radius`	Trace all downstream callers up to 5 hops before refactoring	Free
`query_dependencies`	Forward, reverse, and circular dependency tracing between modules	Free
`check_reachability`	Verify exports are wired to production entry points	Pro
`get_regression_risk`	Score functions by complexity, exposure, churn, and caller count	Pro
`get_unused_code`	Graph-based dead code detection with text-reference backup	Pro
`get_consolidation_opportunities`	Detect duplicate logic, parallel consumers, signature twins	Pro
`get_test_coverage`	Per-module coverage with untested high-complexity function flagging	Pro
`get_vision_docs`	Cross-reference PRDs and specs against implementation	Pro
`get_vision_gaps`	Find specs without code and complex code without specs	Pro
`get_cross_repo_audit`	Compare two repositories for structural duplication and drift	Pro

Example Session

You: "What breaks if I rename formatMessage?"

Pharaoh → get_blast_radius
  Risk: HIGH
  Direct callers: 4 (across 3 modules)
  Transitive impact: 12 functions
  Affected endpoints: POST /api/notifications/send, POST /api/slack/webhook
  Affected cron: daily-digest (09:00 UTC)

You: "Is there already a retry wrapper?"

Pharaoh → search_functions
  Found: withRetry() in src/utils/resilience.ts:42
  Exported: yes | Async: yes | Complexity: 8
  Used by 6 callers across 3 modules
  → Agent imports existing function instead of writing a new one.

How It Works

Tree-sitter parsing (deterministic, no LLM) into a Neo4j knowledge graph
GitHub webhook auto-refreshes on every push to default branch
~60 seconds to map a 50K LOC TypeScript project
Structural metadata only — no source code stored

FAQ

Does Pharaoh store my source code? No. The graph contains function names, file paths, dependency relationships, complexity scores, and export signatures. Never source code.

What languages are supported? TypeScript and Python today. Tree-sitter makes adding languages straightforward.

How does it stay current? GitHub webhook fires on every push. Your graph is always up to date.

Can I use it with private repos? Yes. Read-only access via the GitHub App. Tenant-isolated — your data is never visible to other customers.

How is this different from Sourcegraph / CodeScene / SonarQube? Sourcegraph answers "where is it?" — text search across repos. CodeScene analyzes behavioral patterns from git history. SonarQube does line-level static analysis and linting. Pharaoh answers "what breaks if I change this?" — structural and architectural intelligence built for AI agents, not dashboards.