Power your agent
การติดตั้งจากผู้เผยแพร่ ไม่ใช่เอนด์พอยต์ของ PlugMCP
URL แบบรีโมตหรือคำสั่งบนเครื่องของคุณ? การติดตั้งสองแบบต่างกันอย่างไร
รันบน PlugMCP
A new server you create. It is not this listing's official endpoint. Use the publisher install when you want their URL and their credentials.
เปิดBreak the Claude Pro wall. Offload heavy lifting to Gemini 3.x and GLM-5 via MCP while Claude orchestrates. Reduce consumption by 10x.
Claude Pro users paying $20/month and even those paying $ 100 per month frequently encounter a frustrating reality:
Most power users exhaust their weekly message limits in approximately 2 days. This leaves them blocked for the remaining 5 days of the week. You are effectively paying for a service that is unavailable 71% of the time -- $100/month for 28% availability unable to do heavy work. This creates a scarcity mindset where users hesitate to use Claude for complex tasks, fearing they will run out of limits before the week resets.
Since the release of Claude 4.6, multiple users have reported noticeable performance degradation in Sonnet 4.5. When Sonnet 4.5 is forced to handle orchestration, file I/O, and heavy code generation simultaneously, reasoning quality drops. It begins to lose track of project structure, hallucinate file paths, or forget earlier instructions mid-session.
This MCP approach specifically helps Sonnet 4.5 users by offloading heavy analysis and code generation to capable external models, letting Sonnet focus on what it does best: orchestration and file I/O. The result is superior work output from Sonnet 4.5 compared to making it do everything itself.
Opus 4.6 remains the gold standard for intelligence, but its consumption rate is punishing -- 3-5x higher than Sonnet. On the Pro plan, using Opus for documentation reading, boilerplate generation, or routine analysis burns through your weekly allocation at an alarming rate.
By delegating heavy work to external models, you get Opus-quality orchestration and Parallel Agents with 80% to 90%+ reduction in consumption.
claude-additional-models-mcp is a production-ready Model Context Protocol (MCP) server that integrates external AI models directly into Claude Desktop. Instead of Claude doing everything itself, it delegates heavy computational work to specialized external models and focuses on what it does best: orchestration, file management, and final synthesis.
The result:
While the original GLM-5 (Z.ai) integration offers a massive 744B parameter model with strong reasoning, we observed frequent interruptions and timeouts during long-running delegated tasks. Sessions would break mid-generation, forcing users to retry and wasting both time and tokens.
v2.0 introduces Google Gemini 3.x as the recommended integration:
Both integrations remain fully supported. Choose based on your needs.
The Model Context Protocol allows Claude Desktop to discover and execute tools provided by a local server. This MCP server acts as a bridge: Claude calls a tool (e.g., ask_gemini_pro), the server forwards the request to the external model's API, and returns the result to Claude.
Claude never needs to "think through" the heavy content. It sends the task, receives the output, and writes it to disk.
The Gemini integration uses a dual-path communication strategy for maximum capability:
generativelanguage.googleapis.com/v1beta/openai/): Used for standard chat completions -- ask_gemini and ask_gemini_pro tool calls.generativelanguage.googleapis.com/v1beta): Used for advanced features like search grounding (web_search) and URL context (web_reader). These features are not available through the OpenAI-compatible interface.When Claude Desktop starts, it reads claude_desktop_config.json, launches the MCP server as a child process, and queries it for available tools. The tools then appear in Claude's interface alongside its built-in capabilities.
git clone https://github.com/Arkya-AI/claude-additional-models-mcp.git
cd claude-additional-models-mcp/gemini
npm install
Open your Claude Desktop configuration file:
~/Library/Application Support/Claude/claude_desktop_config.json%APPDATA%\Claude\claude_desktop_config.jsonAdd the Gemini MCP server:
{
"mcpServers": {
"gemini": {
"command": "node",
"args": ["/absolute/path/to/claude-additional-models-mcp/gemini/index.js"],
"env": {
"GOOGLE_API_KEY": "your-api-key-here"
}
}
}
}
Important: Replace
/absolute/path/to/with the actual path where you cloned the repo.
Quit Claude Desktop completely (from the menu bar, not just closing the window) and relaunch it.
ask_gemini, ask_gemini_pro, web_search, web_reader, and parse_documentgit clone https://github.com/Arkya-AI/claude-additional-models-mcp.git
cd claude-additional-models-mcp
npm install
{
"mcpServers": {
"glm5": {
"command": "node",
"args": ["/absolute/path/to/claude-additional-models-mcp/index.js"],
"env": {
"ZAI_API_KEY": "your-api-key-here"
}
}
}
}
Same process as Gemini above. Tools will be named ask_glm5, ask_glm5_pro, etc.
Both integrations expose the same 5-tool interface. The tool names differ by prefix (ask_gemini vs ask_glm5).
To maximize efficiency, Claude is instructed (via CLAUDE.md) to follow this execution hierarchy for every task:
When facing a task with multiple independent parts, Claude spawns sub-agents that each independently call the external model. This is the fastest execution path and must be preferred over sequential calls.
Example: Writing a full-stack feature -- Sub-agent A generates the backend via ask_gemini_pro, Sub-agent B generates the frontend, Sub-agent C generates the tests. All run in parallel.
For tasks that cannot be parallelized, Claude sends the entire prompt directly to the external model.
Example: "Analyze this codebase and explain the authentication flow" -- a single ask_gemini_pro call.
Claude should only perform work internally when:
Claude (Parent)
|-- Planning & coordination (stays in Claude)
|-- File operations & disk I/O (stays in Claude)
|-- Quick responses <100 words (stays in Claude)
|
|-- PRIORITY 1: Spawn parallel sub-agents (for multi-part tasks)
| \-- Each sub-agent uses external model for code/analysis
| |-- ask_[model]_pro for code generation
| |-- ask_[model] for analysis/docs
| \-- Sub-agent writes output to disk
|
\-- PRIORITY 2: Delegate to external model directly (for single-unit tasks)
|-- Analysis >300 words
|-- Code generation >50 lines
|-- Research synthesis from multiple sources
\-- Document processing & OCR
The CLAUDE.md file in this repository acts as a "constitution" for all Claude sessions. It enforces that sub-agents must also delegate to the external model -- they cannot generate large content blocks themselves.
The pattern: External model generates content -> Claude/sub-agent writes output to disk.
Any sub-agent that generates >50 lines of code or >300 words of content without delegating to the external model is violating the delegation rules.
Sonnet 4.5 excels at following complex instructions, managing project structure, and coordinating multi-step workflows. Its degradation occurs when it is forced to also do the heavy lifting.
Result: Sonnet becomes a lean orchestrator. Output quality increases because it is no longer context-overloaded.
Opus should function as the strategic decision-maker, not the laborer.
Result: 90%+ reduction in Opus consumption. You get Opus-level orchestration at Sonnet-level cost.
Four critical rules to ensure delegation is working:
web_search -> web_reader -> ask_gemini synthesizes -> Claude presents the resultparse_document x3 -> ask_gemini comparison -> Claude formats outputask_gemini generates the content -> Claude adapts to the project voice -> writes to diskask_gemini call1. Claude receives: "Research the current state of edge AI deployment"
2. Claude calls web_search("edge AI deployment trends 2025", count=30)
3. Claude calls web_reader() on the top 5 URLs (parallel)
4. Claude calls ask_gemini("Synthesize these sources into a 2000-word report
covering: key trends, major players, deployment challenges, and outlook")
5. Claude writes the report to research-output.md
6. Claude presents a 3-sentence summary to the user
Total Claude consumption: ~800 tokens (orchestration + file I/O)
1. Claude receives: "Build a FastAPI backend with user auth and CRUD endpoints"
2. Claude calls ask_gemini_pro("Generate SQLAlchemy models and Pydantic schemas
for a user authentication system with roles and permissions")
3. Claude writes models.py to disk
4. Claude calls ask_gemini_pro("Generate FastAPI CRUD routes for these models: [models]")
5. Claude writes routes.py to disk
6. Claude runs tests locally
Total Claude consumption: ~500 tokens (orchestration + file I/O)
claude-additional-models-mcp/
├── index.js # GLM-5 MCP server (Z.ai 744B model)
├── package.json # Root dependencies (@modelcontextprotocol/sdk, node-fetch)
├── gemini/
│ ├── index.js # Gemini 3.x MCP server (hybrid API approach)
│ └── package.json # Gemini dependencies (@modelcontextprotocol/sdk, openai)
├── CLAUDE.md # GLM-5 delegation rules (root-level)
├── gemini/CLAUDE.md # Gemini delegation rules
├── EXAMPLES.md # Real-world usage examples and templates
├── CONTRIBUTING.md # Contribution guidelines
├── LICENSE # MIT License
├── .github/ # GitHub issue/PR templates
└── .gitignore
npm scripts:
npm run start:glm5 # Start GLM-5 server
npm run start:gemini # Start Gemini server
claude_desktop_config.json with a JSON linter. A single trailing comma will break it.index.js must be absolute, not relative.node /path/to/gemini/index.js in a terminal. If it errors, the issue is in the server setup.GOOGLE_API_KEY or ZAI_API_KEY is valid and active.max_tokens to 16384 for Gemini or 4000 for GLM-5.GOOGLE_API_KEY for Gemini, ZAI_API_KEY for GLM-5.GLM-5 can experience timeouts during peak hours or with very long prompts. If reliability is critical, switch to the Gemini integration.
Audit checklist:
ask_gemini?web_reader?ask_gemini_pro?parse_document?Expected token patterns per task:
If Claude consumption exceeds 3,000 tokens on analytical work, delegation is not optimized.
Contributions are welcome! Whether it is adding a new model provider, optimizing the sub-agent delegation prompts, or improving error handling, please see CONTRIBUTING.md for guidelines.
This project is licensed under the MIT License. See the LICENSE file for details.
Built by Arkya AI to make Claude truly limitless.
Reduce Claude Desktop consumption by 10x - Integrate Google's Gemini or Z.ai's GLM-5 (744B params) with Claude via MCP for intelligent task delegation
เครื่องมือ
ยังไม่มีรายการเครื่องมือ เซิร์ฟเวอร์แบบรีโมตจะได้รายการหลังการตรวจสอบสถานะครั้งแรก
FAQ
คัดลอกคอนฟิก stdio สำหรับ Cursor, Claude Desktop, Claude Code หรือ VS Code ได้ในหน้านี้ คอนฟิกนี้ใช้การติดตั้งของผู้เผยแพร่ ไม่ใช่พร็อกซีของ PlugMCP
ระเบียนนี้ไม่ได้ระบุตัวแปรสภาพแวดล้อมที่จำเป็น
Claude Additional Models ทำงานแบบโลคัลผ่าน stdio
ตรวจสอบล่าสุดเมื่อ 2026-10-27 Reduce Claude Desktop consumption by 10x - Integrate Google's Gemini or Z.ai's GLM-5 (744B params) with Claude via MCP for intelligent task delegation