Community Developer Tools 10k likes

An MCP server that lets Claude Code, Codex, or Gemini CLI pull in other AI models mid-task, for a second opinion, a multi-model debate, or a full code review, without losing conversation context.

Instant setup via uvx

uvx --from git+https://github.com/BeehiveInnovations/pal-mcp-server.git pal-mcp-server

Source: https://github.com/BeehiveInnovations/zen-mcp-server

About

This server sits between your CLI-based coding agent and a pool of other models (Gemini, OpenAI, Grok, Azure, Ollama, OpenRouter, DIAL) so a single session can hand work to whichever model fits: ask Gemini Pro to sanity-check an architecture call, get GPT-5 and Grok to debate a design tradeoff, or spin up a fresh-context subagent to do a security audit and report back only the conclusion. Conversations thread across tools and models, so switching providers mid-task doesn't mean re-explaining everything.

The project was originally released as Zen MCP and has since been renamed to PAL MCP ('Provider Abstraction Layer') under the same BeehiveInnovations org and repo history: same tool set, same underlying approach, new name. Its core tools cover collaborative brainstorming, extended reasoning, multi-model consensus, and a structured set of code-quality workflows (debug, code review, pre-commit validation) that are deliberately multi-step rather than one-shot, so the model is pushed to investigate before it answers.

Key features

  • Routes a single coding session across Gemini, OpenAI, Grok, Azure, Ollama, OpenRouter, and on-device models
  • clink tool bridges to external CLIs (Gemini CLI, Codex CLI) and can spawn isolated subagents for a task
  • Multi-model consensus tool for getting structured, stance-steered opinions from several models at once
  • Guided, multi-step workflows for debug, code review, and pre-commit validation instead of single-shot answers
  • Conversation threading that survives switching models or tools mid-task
  • Optional tools (analyze, refactor, testgen, secaudit, docgen, tracer) can be enabled individually to save context

Use cases

  • Getting a second model's opinion on a tricky bug before committing to a fix
  • Running a multi-model consensus debate to decide between two architectural approaches
  • Offloading a full code review to an isolated subagent so it doesn't pollute the main session's context
  • Forcing a current-year API/SDK documentation lookup instead of relying on a model's stale training data

Available tools

chat

Brainstorm, get a second opinion, or validate an approach with another model.

clink

Bridges to an external AI CLI (Gemini, Codex) and can launch isolated subagents for a task.

thinkdeep

Extended reasoning pass over edge cases and alternative approaches.

consensus

Collects stance-steered opinions from multiple models on the same question.

debug

Systematic root-cause investigation workflow rather than a single guess.

codereview

Produces a severity-ranked review with actionable feedback.

Frequently asked questions

Do I need API keys for every model it supports?

No. It only activates providers you've given credentials for in your .env; Ollama works with no key at all for local models.

Why is the repo called zen-mcp-server but the README says PAL MCP?

The project was renamed from Zen MCP to PAL MCP; the original repository and history carried over, so the old name still resolves to the current codebase.