Pal MCP Server — Multi-Model AI Gateway for Claude Code

Why multi-model matters inside one agent

Claude Code is excellent at reasoning. Gemini 2.5 Pro has a 2M context window. GPT-4o is fast. Grok has live web access. Ollama runs offline. A real dev team uses all of them. Pal MCP collapses that into one tool call from Claude Code's perspective — ask it to "call Gemini on this 1.5M-token codebase" and Pal routes the request, returns the result, and maintains conversation continuity.

Single-config setup

Add to .mcp.json:

{
  "mcpServers": {
    "pal": {
      "command": "uvx",
      "args": ["--from", "git+https://github.com/BeehiveInnovations/pal-mcp-server.git", "pal-mcp-server"],
      "env": {
        "GEMINI_API_KEY": "your-gemini-key",
        "OPENAI_API_KEY": "your-openai-key",
        "DEFAULT_MODEL": "auto"
      }
    }
  }
}

Restart Claude Code. Now pal_chat, pal_route, and pal_continue are callable.

The routing logic

Set DEFAULT_MODEL=auto and Pal picks a model based on task heuristics:

Override per-call with pal_chat(model="gpt-4o").

CLI-to-CLI bridge

Pal exposes a raw CLI bridge: call Aider, Continue, or any CLI-based agent from within Claude Code. Useful for chaining specialized agents in a single workflow.

Conversation continuity

Every Pal call can continue an existing thread:

pal_continue(thread_id="xyz", prompt="refactor based on Gemini's suggestions")

Thread state is persisted in SQLite under ~/.pal/threads.db. Survives restarts.

Supported providers in 2026

• Anthropic (Claude Opus, Sonnet, Haiku)
• OpenAI (GPT-4o, o3, o3-mini)
• Google (Gemini 2.5 Pro, Flash)
• xAI (Grok-3)
• DeepSeek (R1, V3)
• Ollama (local, 50+ models)
• LiteLLM (proxy for 100+ more)

Cost control

Pal emits a cost-summary per session: total tokens, per-model breakdown, $ estimate. Use MAX_COST_PER_SESSION=5 env var to hard-stop runaway loops.

When Pal is not the right choice

1. Single-model workflows — overhead not worth it, use the provider SDK directly.
2. Production agents — MCP is still evolving; use LiteLLM Proxy for production-grade routing.
3. Compliance-regulated environments — each upstream provider has different data policies; Pal doesn't unify compliance.