PAL MCP (formerly Zen MCP) is a Model Context Protocol server that gives a coding CLI or IDE client, such as Claude Code, Codex CLI, Gemini CLI, Qwen Code, or Cursor, access to other models within a single prompt. The host CLI stays in charge of the workflow while it consults Gemini, OpenAI, Anthropic, Grok, Azure, OpenRouter, DIAL, Ollama, or on-device models for individual subtasks.
Conversation threading carries context across tools and models, so a chain such as a multi-model code review, a plan, an implementation pass, and a pre-commit check happens in one thread where each model sees what earlier models said. The same mechanism lets a session continue after the host's context resets by having another model restate the discussion, and it works around the MCP 25K token limit for large prompts and responses.
The clink tool bridges to external CLIs directly: from inside one CLI you can spawn an isolated instance of another (or the same) CLI as a subagent with a role such as planner or codereviewer. The subagent runs in a fresh context with its own web search, file inspection, and MCP tools, and returns only its final result to the parent session.
Each tool ships as a multi-step workflow with its own parameters, so several tools are disabled by default to save context-window space and can be enabled per install. A provider is activated by adding its credentials to .env.
Features
- Multi-model orchestration: one prompt can involve several providers, with automatic model selection or an explicit choice per subtask
- Conversation continuity: context flows across tools and models within a thread, and can revive a host session after a context reset
- CLI subagents (clink): launch isolated Gemini, Codex, or Claude Code instances with role-specific system prompts and receive only their final report
- Collaboration tools: chat, thinkdeep, planner, and consensus for second opinions, extended reasoning, structured plans, and stance-steered multi-model votes
- Code quality tools: debug for hypothesis-tracked root-cause analysis, codereview with severity levels, precommit validation, and analyze for whole-codebase structure
- Development tools: refactor, testgen, secaudit (OWASP Top 10), and docgen, disabled by default
- Utilities: apilookup forces current-year documentation lookups in a sub-process, challenge prompts critical analysis instead of agreement, tracer maps call flows
- Vision input: screenshots and diagrams are passed to vision-capable models
- Local models: Ollama and other local endpoints run alongside hosted providers
