EX MCP Server

EX MCP Server is a Model Context Protocol (MCP) server that connects modern LLM providers and tools to MCP‑compatible clients (e.g., Claude Desktop/CLI). It provides a unified set of analysis, debugging, refactoring, documentation, testing, and project automation tools accessible over the MCP stdio protocol.

Key Capabilities

• Unified MCP server exposing rich development tools:

  • analyze, codereview, debug, refactor, tracer, testgen, precommit, listmodels, version

• Provider integrations:

  • KIMI (Moonshot), GLM (Zhipu), OpenRouter, and custom OpenAI‑compatible endpoints

Quick Links

• O&M Manual: docs/o_and_m_manual/

• Chat Sheet: docs/o_and_m_manual/CHAT_SHEET.md

• Augment Code Guidelines (EXAI‑MCP default): docs/policies/AUGMENT_CODE_GUIDELINES.md

• Sweep Reports: docs/sweep_reports/

• MCP‑first architecture:

  • Subprocess stdio transport with direct config examples for Claude Desktop/CLI

• Docker and local dev support:

  • Docker image build/publish, local virtualenv (.venv), and cross‑platform scripts

Installation

Prerequisites

• Python 3.9+

• Git

• For local dev: virtualenv support

• Optional: Docker and Docker Compose

Clone

Setup (local)

Run (local)

Configure a client (Claude Desktop/CLI)

Minimal example (stdio):

See the examples/ directory for more configs (macOS, WSL, desktop CLI variants).

Docker

Build and run locally:

A reverse proxy example is provided (nginx.conf) and a remote compose file (docker-compose.remote.yml) that exposes the server as ex-mcp.

Usage Overview

• Use the version tool to verify install:

• Common tools:

  • analyze: smart file analysis

  • codereview: professional code review

  • debug: debugging assistant

  • refactor: code refactoring

  • tracer: static analysis / call chain aid

  • testgen: test generation

  • precommit: quick pre-commit validation

  • listmodels: show available models/providers

Provider‑native Web Browsing Schemas

• Kimi (Moonshot): inject an OpenAI function tool named "web_search" with a string parameter "query".

• GLM (Zhipu): enable tools = [{"type":"web_search","web_search":{}}] only when allowed by env.

• Set these via env for production readiness:

  • KIMI_ENABLE_INTERNET_TOOL=true and KIMI_INTERNET_TOOL_SPEC to a valid JSON tool schema

  • GLM_ENABLE_WEB_BROWSING=true when appropriate (and other GLM browsing flags as documented)

Hidden Model Router (Auto Model Selection)

The server can auto-select a concrete model at the MCP boundary so users don’t need to specify one.

• Enable: HIDDEN_MODEL_ROUTER_ENABLED=true

• Sentinels: ROUTER_SENTINEL_MODELS=glm-4.5-flash,auto

• Default: DEFAULT_MODEL=glm-4.5-flash (a sentinel)

Behavior:

• If a tool requires a model and incoming model is a sentinel (or "auto"), the server resolves a concrete model.

• Structured logs emitted by the server (logger name: "server"):

  • EVENT boundary_model_resolution_attempt input_model=... tool=... sentinel_match=... hidden_router=...

  • EVENT boundary_model_resolved input_model=... resolved_model=... tool=...

Notes:

• The Consensus tool intentionally does not resolve models at the MCP boundary (requires_model = False). You will see the "attempt" log at the boundary, and per-step model selection happens inside the tool.

Tip: Use listmodels to see configured providers/models.

Agentic Audit with Real Models (EX‑AI)

Use a consensus-based, multi-model audit to find issues and get direct fixes.

1. Set provider keys in .env:

• KIMI_API_KEY=...

• GLM_API_KEY=...

2. Run the audit script:

Or rely on env defaults (GLM_AUDIT_MODEL, KIMI_AUDIT_MODEL) and just:

The script returns JSON:

3. Interpreting results:

• Each issue has “direct_fix” with exactly what to change and where.

• Re-run after fixes to validate improvements.

Tests: End-to-end (no real keys required)

We include an “ultimate” test file designed for EX‑AI‑style validation:

• tests/test_e2e_exai_ultimate.py

• Each assert prints a Direct Fix if it fails.

• Run: python -m pytest -q tests/test_e2e_exai_ultimate.py

CI/Test Hygiene (EX fork)

This fork disables some upstream providers by design. If you run the full test suite, import errors may occur for those optional providers. See docs/ci-test-notes.md for ways to skip/guard those tests in CI.

Configuration

• Environment file: .env (see .env.example for available variables)

• Key variables:

  • DEFAULT_MODEL, LOCALE, MAX_MCP_OUTPUT_TOKENS

  • Provider keys: KIMI_API_KEY, GLM_API_KEY, OPENROUTER_API_KEY

  • Custom API: CUSTOM_API_URL, CUSTOM_API_KEY

• Logging: logs/ directory (Docker and local scripts manage ownership/paths)

Attribution

This project is based on the original work at:

• https://github.com/BeehiveInnovations/zen-mcp-server We have forked/copied and adapted it to create EX MCP Server. Attribution to the original authors is preserved.

Our EX‑specific Changes (Zen → EX)

• Rebranding:

  • Service name: zen-mcp → ex-mcp

  • Non-root user: zenuser → exuser (Dockerfile, file ownership)

  • Virtual environment: .zen_venv → .venv

  • Branding strings: “Zen MCP Server” → “EX MCP Server”

• Examples/configs:

  • Server IDs: "zen" → "ex"

  • Commands: zen-mcp-server → ex-mcp-server

  • Paths updated to ex-mcp-server

• CI/workflows & templates:

  • GitHub discussions/links point to ex-mcp-server

  • GHCR image names: ghcr.io//ex-mcp-server:...

• Architecture intent:

  • MCP-first stdio transport, reverse proxy alignment, and consistent service naming

Contributing

Please see CONTRIBUTING.md for development workflow, coding standards, and testing.

License

See the LICENSE file in this repository.

Additional Resources

• MCP Spec: https://modelcontextprotocol.io/

• Claude Desktop docs for MCP: https://docs.anthropic.com/claude/docs/model-context-protocol

• Original source (upstream): https://github.com/BeehiveInnovations/zen-mcp-server

• Current project: https://github.com/BeehiveInnovations/ex-mcp-server