Zen MCP Server — Use Cases, Install & Live Demo

Why use it

Key features

Cross-model conversations: each tool can pick a different model
Workflow tools: thinkdeep, planner, debug, codereview, consensus, refactor
Provider support: Gemini, OpenAI, Grok, OpenRouter, Ollama, Azure OpenAI
Continuation IDs to keep cross-model context alive
Built-in cost tracking per provider

Live Demo

What it looks like in practice

zen-mcp-server.replay ▶ ready

0/0

Install

Pick your client

~/Library/Application Support/Claude/claude_desktop_config.json · Windows: %APPDATA%\Claude\claude_desktop_config.json

{
  "mcpServers": {
    "zen-mcp-server": {
      "command": "uvx",
      "args": [
        "zen-mcp-server"
      ]
    }
  }
}

Open Claude Desktop → Settings → Developer → Edit Config. Restart after saving.

~/.cursor/mcp.json · .cursor/mcp.json

{
  "mcpServers": {
    "zen-mcp-server": {
      "command": "uvx",
      "args": [
        "zen-mcp-server"
      ]
    }
  }
}

Cursor uses the same mcpServers schema as Claude Desktop. Project config wins over global.

VS Code → Cline → MCP Servers → Edit

{
  "mcpServers": {
    "zen-mcp-server": {
      "command": "uvx",
      "args": [
        "zen-mcp-server"
      ]
    }
  }
}

Click the MCP Servers icon in the Cline sidebar, then "Edit Configuration".

~/.codeium/windsurf/mcp_config.json

{
  "mcpServers": {
    "zen-mcp-server": {
      "command": "uvx",
      "args": [
        "zen-mcp-server"
      ]
    }
  }
}

Same shape as Claude Desktop. Restart Windsurf to pick up changes.

~/.continue/config.json

{
  "mcpServers": [
    {
      "name": "zen-mcp-server",
      "command": "uvx",
      "args": [
        "zen-mcp-server"
      ]
    }
  ]
}

Continue uses an array of server objects rather than a map.

~/.config/zed/settings.json

{
  "context_servers": {
    "zen-mcp-server": {
      "command": {
        "path": "uvx",
        "args": [
          "zen-mcp-server"
        ]
      }
    }
  }
}

Add to context_servers. Zed hot-reloads on save.

claude mcp add zen-mcp-server -- uvx zen-mcp-server

One-liner. Verify with claude mcp list. Remove with claude mcp remove.

Use Cases

Real-world ways to use Zen MCP Server

Get consensus from 3 different models on a hard architectural decision

👤 Tech leads making high-stakes calls ⏱ ~20 min intermediate

When to use: Pure Claude opinions feel like a single point of view; you want diversity.

Prerequisites

API keys for at least 2 alt providers — Set GEMINI_API_KEY, OPENAI_API_KEY, OPENROUTER_API_KEY

Flow

Frame the question

Use zen.consensus. Question: should we move our queue from Redis to NATS Jetstream? Stakes: 50M msgs/day, 99.99% uptime. Ask Gemini Pro and o3.✓ Copied

→ Both models respond, Claude summarizes diff between answers
Drill into the disagreement

Where do they disagree? Use zen.chat to push back on each model's weakest point.✓ Copied

→ Per-model rebuttal pass
Synthesize

Give me a final recommendation with the strongest argument from each side surfaced.✓ Copied

→ Merged decision doc

Outcome: An architectural call you can defend with cross-model evidence.

Pitfalls

Models hallucinate API quirks differently — false consensus — Always verify with the actual docs/source for fact claims

Combine with: context7 · git-mcp-idosal

Use thinkdeep with Gemini Pro to plan before Claude executes

👤 Devs who hit Claude's planning ceiling on huge tasks ⏱ ~40 min intermediate

When to use: Task is so big that even extended thinking loses the thread.

Flow

Hand off the brief

Use zen.thinkdeep with gemini-2.5-pro. Plan the migration from monolith to 3 microservices. Identify all sequencing risks.✓ Copied

→ Long plan from Gemini, returned as structured doc
Critique the plan

Now you (Claude) review that plan. What's missing? What's risky?✓ Copied

→ Gap analysis
Execute step 1

Implement the first phase of the plan in this repo.✓ Copied

→ Real edits

Outcome: Plans you trust because two models agreed on shape.

Combine with: desktop-commander-mcp

Run zen.codereview from a different model than the one that wrote the code

👤 Solo devs who want adversarial review ⏱ ~25 min intermediate

When to use: You vibed a feature with Claude and want a fresh pair of eyes (other model).

Flow

Stage the diff

Save the current branch's diff. Pass it to zen.codereview with o3.✓ Copied

→ Review comments organized by file
Apply or reject

Show me each suggestion. Apply only the ones I approve.✓ Copied

→ Per-suggestion accept/reject prompt

Outcome: Code reviewed by a model with different priors.

Combine with: git-mcp-idosal

Combinations

Pair with other MCPs for X10 leverage

zen-mcp-server + desktop-commander-mcp

Plan with Gemini, execute with Claude+desktop-commander

thinkdeep the migration → desktop-commander to apply step 1✓ Copied

zen-mcp-server + git-mcp-idosal

Pull repo via gitmcp, codereview via zen with o3

Codereview the PR diff from owner/repo using o3 — second opinion.✓ Copied

Tools

What this MCP exposes

Tool	Inputs	When to call	Cost
chat	prompt, model?, continuation_id?	General Q&A to alt model	1 API call to chosen model
thinkdeep	problem, model?	Planning hard problems	Heavy — uses thinking budget
codereview	files[], model?	Fresh-eyes PR review	1 API call
debug	error, context, model?	Stuck on a bug Claude can't see	1 API call
consensus	question, models[]	High-stakes decisions	N API calls
precommit	diff, model?	Final check before commit	1 API call

Cost & Limits

What this costs to run

API quota: Per-provider; you pay each one separately
Tokens per call: Variable by tool — thinkdeep can hit 30k+
Monetary: Free server; pay each LLM provider directly
Tip: Set DEFAULT_MODEL=ollama:llama3.1 for casual chat, escalate to paid only when needed

Security

Permissions, secrets, blast radius

Minimum scopes: outbound:llm-providers

Credential storage: Provider keys in env vars

Data egress: Whichever provider(s) you configure

Sending your code to multiple providers means multiple data-handling policies — review each provider's terms

Troubleshooting

Common errors and fixes

Model not available

Check the env var for that provider; only configured providers show up

Verify: List zen tools — model defaults appear there

Continuation ID lost

Each tool returns a continuation_id; pass it back to keep cross-model context

Cost runaway on thinkdeep

Set MAX_THINKING_TOKENS in env; default can be high

Alternatives

Zen MCP Server vs others

Alternative	When to use it instead	Tradeoff
context-mode-mcp	You want context shaping, not multi-model routing	Different problem
Direct OpenAI/Gemini MCPs	You only need one alt model, not orchestration	Less ceremony but no consensus/thinkdeep workflows

More

Resources

📖 Read the official README on GitHub

🐙 Browse open issues

🔍 Browse all 400+ MCP servers and Skills