PAL
PAL (formerly Zen MCP) lets Claude Code and other CLI agents call out to other AI models mid-conversation for a second opinion, a deep review, or a full multi-model debate.
Clone and run
git clone https://github.com/BeehiveInnovations/pal-mcp-server.git && cd pal-mcp-server && ./run-server.sh
Source: https://github.com/BeehiveInnovations/pal-mcp-server
About
PAL MCP Server (a Provider Abstraction Layer, renamed from its earlier Zen MCP identity) plugs into Claude Code, Codex CLI, Gemini CLI, and similar tools so they can hand parts of a task to other models (Gemini, OpenAI, Grok, Azure, OpenRouter, DIAL, or local Ollama models) without losing the original conversation's context. A 'context revival' mechanism lets a conversation resume with another model even after the original session's context window has reset.
Its tool set is opinionated about workflow, not just chat: codereview and precommit run structured, severity-ranked reviews; debug drives systematic root-cause analysis instead of guesswork; consensus polls multiple models and has them debate a question before answering; and the newer clink tool bridges out to external AI CLIs directly, letting you launch an isolated CLI subagent for a heavy task (like a full codebase audit) so it doesn't pollute your main session's context.
Key features
- Routes tasks to 50+ models across Gemini, OpenAI, Azure, X.AI, OpenRouter, DIAL, and Ollama
- Context revival lets a conversation resume with a different model after a context reset
- codereview and precommit tools produce severity-ranked findings rather than free-form comments
- consensus tool runs multi-model debate before settling on an answer
- clink bridges to external AI CLIs, running heavy subagent tasks outside the main context window
- Supports local models for zero-cost, fully private analysis
Use cases
- Have Claude Code hand off a gnarly bug to a debug-focused pass before touching any code
- Get a second model's opinion on a design decision via consensus before committing to an approach
- Run a full codereview pass with severity levels before opening a pull request
- Offload a large, context-heavy audit to an isolated CLI subagent via clink so the main session stays clean
Available tools
chat
Runs a multi-turn conversation or brainstorming session with a chosen model.
debug
Performs systematic root-cause analysis on a bug rather than ad hoc guessing.
codereview
Produces a structured code review with severity-ranked findings.
consensus
Gathers opinions from multiple models and has them debate before returning an answer.
clink
Bridges to an external AI CLI, letting a task run in an isolated subagent instance.
planner
Breaks a complex project down into actionable, ordered steps.
Frequently asked questions
Does using PAL mean giving up Claude Code's context when switching models?
No. PAL's context revival is built to let a conversation resume with a different model even after the original context has reset.
What's the difference between PAL and Zen MCP?
They're the same project. PAL is the current name after BeehiveInnovations renamed Zen MCP to avoid confusion with an unrelated product and better reflect its role as a provider abstraction layer.