Labsco
MCP SERVER

Bring several models into one conversation — code review with Gemini, a second opinion from GPT, consensus between them — with context carried across tools.

Model Routing, Multi-Model Consultation & Cost Control
Summary
Six tools ship disabled on purpose, and that is the right default.

Every tool's workflow description occupies context whether or not you call it, so `analyze`, `refactor`, `testgen`, `secaudit`, `docgen` and `tracer` are left out of `DISABLED_TOOLS` until you want them — a one-line change. `version` and `listmodels` cannot be disabled. Cost lands on your own provider keys, not on the client, so a multi-model consensus run is billed by whoever you called.

What it is

A provider abstraction layer for coding CLIs. It lets Claude Code, Codex CLI, Gemini CLI, Cursor and similar clients call other models mid-task and keeps a conversation thread across those calls, so a model brought in at step 11 knows what a different model said at step 7. It ships as `pal-mcp-server`.

What you get
  • A bridge to external AI CLIs, including spawning isolated subagents in a fresh context that return only their final result — `clink`
  • Collaboration and planning tools enabled by default: brainstorming and second opinions, extended reasoning, structured plans, and multi-model debate with stance steering — `chat`, `thinkdeep`, `planner`, `consensus`
  • Code quality work: systematic root-cause investigation, pre-commit validation, and reviews with severity levels — `debug`, `precommit`, `codereview`
  • Utilities that are on by default: forced current-year API and SDK lookups in a subprocess, and a critical-analysis check against reflexive agreement — `apilookup`, `challenge`
  • Off by default, enabled by editing `DISABLED_TOOLS`: architecture and dependency analysis, refactoring, test generation, OWASP Top 10 security audits, documentation generation and call-flow tracing — `analyze`, `refactor`, `testgen`, `secaudit`, `docgen`, `tracer`
  • Conversation continuity across tools and models, including continuing after your CLI's own context resets
  • Automatic workarounds for MCP's 25K limit on large prompts and responses
Requirements

API credentials in `.env` for whichever providers you want — Gemini, OpenAI, Azure, X.AI, OpenRouter, DIAL, Ollama; it activates each provider that has credentials. Two install paths: clone and run `./run-server.sh`, which configures Claude Desktop, Claude Code, Gemini CLI, Codex CLI and Qwen CLI for you, or launch through uvx straight from the repository. `DEFAULT_MODEL`, `DISABLED_TOOLS`, `CONVERSATION_TIMEOUT_HOURS` and `MAX_CONVERSATION_TURNS` are set in the client config. Version 9.8.2.

Setup effort

One command plus a key — uvx --from git+https://github.com/BeehiveInnovations/pal-mcp-server.git pal-mcp-server, then supply credentials