代码知识图谱MCP代理

Pharaoh MCP代理将代码库映射为可查询的知识图谱,使AI编码助手在无头环境中获取架构信息。

  • by Pharaoh-so
  • TypeScript
  • 混合部署
  • 2 星标
  • 更新于
在 GitHub 查看 Pharaoh-so/pharaoh-mcp

使用说明

@pharaoh-so/mcp

MCP proxy for Pharaoh — maps codebases into queryable knowledge graphs for AI agents.

Pharaoh gives AI coding assistants a complete architectural map of your codebase: every function, dependency, module, and connection. Instead of reading files one at a time, your AI agent queries the knowledge graph and gets instant answers about blast radius, unused code, dependency chains, and more.

This package enables Claude Code to connect to Pharaoh in headless environments (VPS, SSH, containers, CI) where a browser isn't available for OAuth. It acts as a stdio-to-SSE proxy, presenting itself as a local MCP server while relaying all communication to the remote Pharaoh server.

Quick Start

Step 1 — Authenticate

Run the proxy directly to trigger the device authorization flow:

npx @pharaoh-so/mcp

This displays a device code and a URL. Open the URL on any device (phone, laptop, tablet) and enter the code to authorize. Credentials are saved to ~/.pharaoh/credentials.json and remain valid for 7 days, with automatic re-authorization when they expire.

Step 2 — Add to Claude Code

npx @pharaoh-so/mcp

Verify the connection:

claude mcp list

You should see pharaoh listed as a stdio server.

Switching from SSE

If you previously added Pharaoh as an SSE server, remove it first:

claude mcp remove pharaoh
npx @pharaoh-so/mcp

How It Works

Claude Code ← stdio → @pharaoh-so/mcp ← SSE/HTTP → mcp.pharaoh.so

The proxy implements the Model Context Protocol (MCP) specification:

  1. Claude Code launches the proxy as a child process and communicates via stdio (stdin/stdout)
  2. The proxy authenticates with Pharaoh using stored credentials (or triggers device flow)
  3. All MCP messages (tool calls, responses, notifications) are relayed to the remote Pharaoh server over SSE (Server-Sent Events)
  4. Pharaoh queries the knowledge graph and returns architectural data to the proxy
  5. The proxy forwards responses back to Claude Code via stdio

Authentication uses RFC 8628 (OAuth 2.0 Device Authorization Grant) — no browser is needed on the machine running Claude Code.

Available Tools

Once connected, Pharaoh provides these MCP tools:

Orient

Tool What it answers
get_codebase_map What modules exist and how do they relate?
get_module_context What does this module look like before I modify it?
search_functions Does this function already exist somewhere?
pharaoh_recon Map, module context, search and blast radius in one call

Investigate

Tool What it answers
get_blast_radius What breaks if I change this function/file/module?

Manage

Tool What it answers
pharaoh_account Check plan, toggle PR Guard, trigger refresh
pharaoh_admin Org-level administration

Inspect Mode

Use --inspect to dump the full tool manifest as JSON (useful for MCP registry validation and debugging):

npx @pharaoh-so/mcp --inspect

This outputs the complete list of tools with their schemas and exits immediately, without connecting to the server.

CLI Options

Usage: pharaoh-mcp [options]

Options:
  --server <url>   Pharaoh server URL (default: https://mcp.pharaoh.so)
  --logout         Clear stored credentials and exit
  --inspect        Output tool manifest as JSON and exit
  --help           Show help
  --version        Show version number

Configuration

Credentials

Credentials are stored at ~/.pharaoh/credentials.json with 0600 permissions (owner-read/write only). The file contains:

  • Access token — used to authenticate MCP requests
  • Refresh token — used to obtain new access tokens when they expire
  • Expiry timestamp — tokens are refreshed automatically before expiration

To clear credentials:

npx @pharaoh-so/mcp --logout

Custom Server

For self-hosted Pharaoh instances or development, register manually, pinned to a version (npx @pharaoh-so/mcp --version prints the one you have):

claude mcp add --scope user pharaoh -- npx @pharaoh-so/mcp@<version> --server https://your-pharaoh-instance.com

Environment Variables

Variable Default Description
PHARAOH_SERVER_URL https://mcp.pharaoh.so Pharaoh server URL (alternative to --server)

Requirements

  • Node.js >= 18
  • Claude Code or any MCP-compatible AI client
  • A Pharaoh account — sign up at pharaoh.so

Security

  • Credentials are stored with restrictive file permissions (0600 — owner-read/write only)
  • Authentication uses the RFC 8628 device authorization flow — no secrets are embedded in the package
  • All communication with the Pharaoh server uses HTTPS/TLS
  • The proxy never reads or sends your local files. Pharaoh reads your code from GitHub through its GitHub App, maps it into a knowledge graph of structural metadata (function names, file paths, dependency relationships), and deletes the clone when mapping ends.
  • Tokens expire after 7 days and are refreshed automatically
  • Only use trusted server URLs — the proxy sends your auth token to the configured server
  • No telemetry. The proxy's only network calls are the device flow, token refresh, the MCP connection, and one registry lookup during interactive setup (below).

What the installer writes on your machine

npx @pharaoh-so/mcp in a terminal is the interactive setup. Every later proxy start (Claude Code launching the server) repeats the plugin, skill and command steps so a new version takes effect, and nothing else.

Path Mode When What
~/.pharaoh/ 0700, set again on every start Always Credentials and the Claude Code plugin marketplace
~/.pharaoh/credentials.json 0600 After sign-in Access token, refresh token, expiry
~/.pharaoh/claude-marketplace/ — ~/.claude/ exists The one-plugin marketplace Claude Code installs as pharaoh@pharaoh: the skills and the Grep routing hook. Rewritten when the package version changes
~/.claude/commands/plan.md, review.md — ~/.claude/ exists Slash commands. A file is written only when it is absent or is Pharaoh's: it carries the front-matter comment # managed-by: pharaoh-mcp, or has the exact header and $ARGUMENTS footer of a Pharaoh command from 0.3.46 or earlier. Any other file with that name is yours: it is kept, a symlink is never followed, and the setup output names it. Rename your file to keep both. 0.3.51 and earlier also wrote sessions.md; later versions delete it while it carries the marker
~/.claude.json — Interactive setup Removes earlier pharaoh entries, then claude mcp add --scope user pharaoh -- npx @pharaoh-so/mcp@<version>
~/.claude/settings.json — Interactive setup, and only if the file already exists Adds to permissions.allow, each by full name: mcp__pharaoh__get_codebase_map, mcp__pharaoh__get_module_context, mcp__pharaoh__search_functions, mcp__pharaoh__get_blast_radius, mcp__pharaoh__get_unused_code, mcp__pharaoh__pharaoh_recon, mcp__pharaoh__pharaoh_account. The setup prints a line saying what it did. The proxy never creates this file and never touches it on a start
~/.agents/skills/pharaoh-review/ — ~/.codex/ exists The review skill for Codex ($pharaoh-review), rewritten on every start. Other tools share ~/.agents/skills/, so nothing else there is touched
~/.openclaw/skills/, ~/.openclaw/openclaw.json — ~/.openclaw/ exists Skills prefixed pharaoh-, and an mcpServers.pharaoh entry running npx @pharaoh-so/mcp@<version>. An entry that is not an npx @pharaoh-so/mcp command is left alone

Versions 0.3.46 and earlier wrote one rule, mcp__pharaoh, on every proxy start. That prefix auto-approved every Pharaoh tool, including account deletion and pharaoh_admin, so a prompt injection could run them without a human prompt. The next interactive setup replaces that rule with the list above; account deletion, pharaoh_admin and any tool not listed prompt before they run. To take the blanket rule out by hand, delete mcp__pharaoh from permissions.allow.

Pinned version and upgrades

Every registration the setup writes runs the exact version that wrote it (npx @pharaoh-so/mcp@<version>). A new release on npm changes nothing on your machine until you upgrade:

npx @pharaoh-so/mcp@latest

This signs you in again, rewrites the registration with the new version and refreshes the plugin. The interactive setup asks the npm registry for the newest version and prints a one-line hint when there is one; proxy starts never check.

Reporting Vulnerabilities

If you discover a security vulnerability, please report it responsibly by emailing security@pharaoh.so. Do not open a public issue.

Troubleshooting

"Connection refused" or "ECONNREFUSED"

The Pharaoh server may be temporarily unavailable. Check status.pharaoh.so or try again in a few minutes.

"Token expired" or "401 Unauthorized"

Re-authenticate by running the proxy directly:

npx @pharaoh-so/mcp

Or clear credentials and start fresh:

npx @pharaoh-so/mcp --logout
npx @pharaoh-so/mcp

"pharaoh" not showing in claude mcp list

Make sure you added it with the correct command:

npx @pharaoh-so/mcp

Note the -- separator between pharaoh and npx.

Device code not working

  • Ensure you're opening the URL on a device with browser access
  • The device code expires after 15 minutes — request a new one by re-running the command
  • Check that you're logged into GitHub when authorizing

Slow startup

The first run after installation may take a moment as npm downloads the package. Subsequent runs use the cached version. To pre-install globally:

npm install -g @pharaoh-so/mcp
claude mcp add --scope user pharaoh -- pharaoh-mcp

How Pharaoh Works

Pharaoh parses your repositories using tree-sitter and maps structural metadata into a Neo4j knowledge graph. The graph contains:

  • Functions — names, signatures, complexity scores, export visibility
  • Files — paths, module membership, language classification
  • Modules — logical groupings detected from directory structure
  • Dependencies — import/export relationships, call chains, module connections
  • Vision specs — PRD/spec documents linked to implementation

No source code is stored — only structural metadata. When an AI agent queries Pharaoh, it gets architectural facts in minimal tokens, not raw code dumps. This means your AI assistant can understand your entire codebase architecture without consuming its context window reading files one by one.

Supported Languages

  • TypeScript / JavaScript — full support (functions, classes, imports, exports, JSX)
  • Python — full support (functions, classes, imports, decorators)
  • More languages planned via tree-sitter grammar support

Contributing

Contributions are welcome. Please open an issue first to discuss what you'd like to change.

  1. Fork the repository
  2. Create your feature branch (git checkout -b feature/amazing-feature)
  3. Commit your changes (git commit -m 'Add amazing feature')
  4. Push to the branch (git push origin feature/amazing-feature)
  5. Open a Pull Request

License

MIT -- Pharaoh, Inc.

Links