Zen MCP Server (now PAL)
The renamed zen-mcp-server: routes Claude Code, Codex or Cursor to Gemini, GPT, Grok and local models in one context.
Add to your client
Copy the config for your MCP client and paste it into its config file.
git clone https://github.com/BeehiveInnovations/pal-mcp-server.git && cd pal-mcp-server && ./run-server.shPaste into ~/Library/Application Support/Claude/claude_desktop_config.json
{
"mcpServers": {
"zen-mcp-server-now-pal": {
"command": "uvx",
"args": [
"--from",
"git+https://github.com/BeehiveInnovations/pal-mcp-server.git",
"pal-mcp-server"
],
"env": {
"GEMINI_API_KEY": "<your-gemini-key>",
"DEFAULT_MODEL": "auto",
"DISABLED_TOOLS": "analyze,refactor,testgen,secaudit,docgen,tracer"
}
}
}
}Requires `uv` (the Python package runner). Install it from https://docs.astral.sh/uv/ if `uvx` is not found.
Step-by-step guides: Add to Claude Desktop · Add to Cursor · Add to Windsurf
Before you start
- Python 3.10+, Git, and uv/uvx installed
- At least one provider API key: OpenRouter, Gemini, OpenAI, Azure OpenAI, X.AI, DIAL — or a local Ollama for free/private models
- A CLI or IDE client to orchestrate from: Claude Code, Codex CLI, Gemini CLI, Qwen Code, Cursor, or the Claude VS Code extension
About Zen MCP Server (now PAL)
First, the name: if you're searching for "zen mcp server", this is it. BeehiveInnovations renamed Zen MCP to PAL MCP (Provider Abstraction Layer) in late 2025 — the old GitHub URL 301-redirects to pal-mcp-server, the docs include a dedicated name-change page, and existing setups only need run-server.sh re-run and any ZEN-prefixed .env vars renamed to PAL. Same 11k+-star project, same tools.
The core idea is conversation threading across models. Your CLI (Claude Code, Codex, Gemini CLI, Qwen, Cursor) stays the orchestrator, and PAL's tools hand subtasks to other models with context intact: codereview can run multi-pass reviews where Gemini Pro sees what O3 flagged ten steps earlier, consensus stages structured debates with stance steering, and the "context revival" trick lets another model re-brief your CLI after its own context window resets. It also works around MCP's 25K token limit for large prompts and responses.
The 16 tools split into enabled-by-default essentials (chat, thinkdeep, planner, consensus, codereview, precommit, debug, apilookup, challenge, plus the clink CLI-to-CLI bridge that can spawn isolated Codex/Gemini subagents) and disabled-by-default extras (analyze, refactor, testgen, secaudit, docgen, tracer) — each enabled tool costs context window in your client, which is why the default DISABLED_TOOLS list exists. Providers activate from whatever keys are present: Gemini, OpenAI, Azure OpenAI, X.AI/Grok, OpenRouter, DIAL, or local Ollama.
Setup is either clone-plus-./run-server.sh (auto-configures Claude Desktop/Code, Gemini, Codex and Qwen CLIs) or the README's instant uvx config that runs straight from the git repo — the uvx invocation is uvx --from git+https://github.com/BeehiveInnovations/pal-mcp-server.git pal-mcp-server, with the README wrapping it in a bash PATH-hunting one-liner for clients that can't find uvx. Trade-offs: you're paying per-token for every consulted model, multi-model workflows add latency, and the biggest wins need at least one strong second model (Gemini 3.0 Pro or GPT-5-class) worth consulting.
Tools & capabilities (16)
clinkCLI-to-CLI bridge: spawn Gemini/Codex/Claude subagents in isolated contexts with role prompts
chatBrainstorm and get second opinions from other models
thinkdeepExtended reasoning and edge-case analysis
plannerBreak complex projects into structured, actionable plans
consensusStructured multi-model debate with stance steering
debugSystematic root-cause investigation with hypothesis tracking
precommitValidate changes before committing to prevent regressions
codereviewMulti-pass reviews with severity levels and actionable feedback
analyzeArchitecture and dependency analysis across codebases (disabled by default)
refactorRefactoring with decomposition focus (disabled by default)
testgenTest generation with edge cases (disabled by default)
secauditOWASP Top 10 security audits (disabled by default)
docgenDocumentation generation with complexity analysis (disabled by default)
apilookupForce current-year API/SDK doc lookups in a sub-process
challengePush back on reflexive agreement with critical analysis
tracerStatic-analysis prompts for call-flow mapping (disabled by default)
What this server can do
Zen MCP Server (now PAL) provides tools for these capabilities — tap one to see every MCP server that does the same:
When to use it
- Multi-model code review: Claude reviews, then Gemini Pro and GPT-5 deep-dive the same findings in one thread
- Consensus decisions: stage a structured debate between models on an architecture choice before committing
- Context revival: after your CLI's context resets, have another model re-brief it and continue the task
- Subagent offloading via clink: spawn an isolated Codex code-review agent that returns only the final report
Quick setup
- 1Get at least one API key (OpenRouter is the single-key way to reach many models; Ollama works for local/free)
- 2Option A: `git clone https://github.com/BeehiveInnovations/pal-mcp-server.git && cd pal-mcp-server && ./run-server.sh` — auto-configures Claude Desktop/Code, Gemini, Codex and Qwen CLIs
- 3Option B: add the uvx config to ~/.claude/settings.json or .mcp.json — `uvx --from git+https://github.com/BeehiveInnovations/pal-mcp-server.git pal-mcp-server` with your API keys in env
- 4Keep the default DISABLED_TOOLS list unless you need analyze/refactor/testgen etc. — every enabled tool consumes context window
- 5Migrating from zen-mcp-server: re-run ./run-server.sh and rename any ZEN-prefixed .env vars to PAL
Security notes
Runs locally over stdio; your provider API keys (Gemini/OpenAI/OpenRouter/etc.) live in env or .env and are sent only to those providers. Be aware that code and files you route through its tools are forwarded to whichever external models you invoke — use the Ollama local-model path if that's unacceptable. Old zen-mcp-server configs keep working via GitHub's redirect, but ZEN-prefixed env vars should become PAL.
Zen MCP Server (now PAL) FAQ
Is Zen MCP Server dead? Why does GitHub redirect to pal-mcp-server?
It's not dead — it was renamed. Zen MCP is now PAL MCP (Provider Abstraction Layer) under the same BeehiveInnovations org; GitHub 301-redirects the old URL, and the maintainers state the software and workflows are identical. The rename avoided confusion with a similarly named product. Update ZEN-prefixed env vars to PAL and re-run run-server.sh.
Is it free and safe to use?
The server is free, Apache-2.0, and runs locally over stdio — but the models it consults aren't free: you pay per-token on every provider you configure (Ollama being the local zero-cost exception). Safety-wise, code you route through its tools is sent to those external model providers, so keep secrets out of prompts or stay on local models.
Do I need multiple API keys for it to be useful?
No — one key is enough to start, and OpenRouter alone unlocks dozens of models through a single key. The value comes from having at least one strong second model (Gemini 3.0 Pro or a GPT-5-class model) for your primary CLI to consult; add more providers only when you want specific strengths like O3 reasoning or local privacy.
Alternatives to Zen MCP Server (now PAL)
Compare all alternatives →
Official MCP server providing persistent, file-backed knowledge-graph memory across sessions.
Structured step-by-step reasoning tool for breaking problems into revisable thought sequences.
Mem0's local-first memory layer: a Dockerized MCP server plus dashboard that keeps agent memories on your machine.
Compare Zen MCP Server (now PAL) with: