Skip to main content
Glama

Switchback MCP

MCP server for Switchback — classify AI agent turns by complexity and pick the cheapest model that can handle them. Drop-in for Claude Desktop, Cursor, Cline, and any MCP client.

npm Powers VibeKit

Built by VibeKit — same cascade router that powers every "Auto" turn in our hosted product. Learn more →

What it does

Exposes two tools over the Model Context Protocol:

  • classify_turn — given a user message, returns {tier: 0|1|2, why, fallback, durationMs}. Use it when you're deciding whether to spend on a flagship model or stay on a cheap one for the next agent step.

  • recommend_model — given a user message + a ladder of models (cheap → flagship), returns the cheapest model on the ladder that can plausibly handle it. Convenience wrapper around classify_turn.

Both tools fire a single small-model call via OpenRouter (default: openai/gpt-5.4-mini, ~$0.0001/call). Fail-soft: any error returns tier 1 with fallback: true rather than blocking your agent.

Related MCP server: oracle-models

Install

npm install -g vibekit-switchback-mcp

Configure

Set your OpenRouter key (get one at https://openrouter.ai/keys):

export OPENROUTER_API_KEY=sk-or-...

Optional: override the classifier model:

export SWITCHBACK_CLASSIFIER_MODEL=anthropic/claude-haiku-4.5

Claude Desktop

Add to your ~/Library/Application Support/Claude/claude_desktop_config.json:

{
  "mcpServers": {
    "switchback": {
      "command": "npx",
      "args": ["-y", "vibekit-switchback-mcp"],
      "env": {
        "OPENROUTER_API_KEY": "sk-or-..."
      }
    }
  }
}

Cursor / Cline / others

Any MCP client that supports stdio servers. Same npx invocation.

Example calls

classify_turn:

{
  "userMessage": "rebuild the auth flow with passkeys"
}

returns:

{
  "tier": 2,
  "why": "architectural rework + new auth method",
  "fallback": false,
  "durationMs": 412
}

recommend_model:

{
  "userMessage": "fix typo in README",
  "ladder": [
    "openai/gpt-5.4-mini",
    "openai/gpt-5.4",
    "openai/gpt-5.5"
  ]
}

returns:

{
  "recommended_model": "openai/gpt-5.4-mini",
  "tier": 0,
  "ladder_size": 3,
  "classifier": { "tier": 0, "why": "trivial copy edit", "fallback": false, "durationMs": 318 }
}

When to use it

  • Multi-step agents where you can swap models between rounds.

  • BYOK-style products where you want to keep all routing inside the user's chosen brand — pass a brand-locked ladder per user.

  • Cost-conscious orchestrators that want to avoid spending flagship rates on trivial turns.

When NOT to use it

  • One-shot chat completions where you can't change models after the first call — use OpenRouter's auto or NotDiamond/Martian instead.

  • Latency-critical sub-200ms-first-token UX — the classifier adds 200-500ms.

License

MIT. See LICENSE.

Built by VibeKit. The core library is vibekit-switchback — this package is the MCP wrapper.

Available Tools

2 tools
classify_turnA

Classify how complex a user request is on a 0/1/2 scale. Use this when you're orchestrating an agent and want to decide whether to spend on a flagship model (tier 2), a mid-tier model (tier 1), or stay cheap (tier 0). Returns {tier, why, fallback, durationMs}.

Inputs: userMessage (required), recentHistory.

recommend_modelA

Given a ladder of models (cheap → flagship) and a user message, return the cheapest model on the ladder that can plausibly handle it, plus reasoning. The ladder should be brand-locked (all-Anthropic, all-OpenAI, etc.) to honor your user's BYOK provider.

Inputs: ladder (required), userMessage (required), recentHistory.

Tool Schema Changelog

Recent tool additions, removals, and schema changes observed during successful MCP inspections.

Loading tool schema changelog…

    TDQS

    A3.8/5.0

    Scored across 2 tools

    Disambiguation3/5

    The two tools both relate to routing user requests to models, so an agent could confuse classifying a turn with recommending a model. However, the inputs and outputs are distinct enough—one returns a tier, the other returns a specific model—to be workable with clear descriptions.

    Naming Consistency5/5

    Both tools follow a consistent verb_noun snake_case pattern: classify_turn and recommend_model. Naming is predictable and clearly indicates the action and object.

    Tool Count3/5

    With only two tools, the server feels thin, but the scope is narrow enough that both tools can earn their place. The count is borderline rather than egregiously insufficient.

    Completeness4/5

    The server covers the core routing workflow: classifying complexity and selecting a cost-appropriate model. Minor gaps exist, such as no combined route tool or ladder management, but agents can work around these using the provided inputs.

    Maintenance

    ActivityMaintained
    ResponsivenessOutdated data

    Related MCP Connectors

    Related MCP Servers

    • A
      license
      A
      quality
      D
      maintenance
      Classifies development task complexity (LIGHT/MEDIUM/HEAVY) and recommends the most cost-efficient AI model per provider, enabling optimized model selection for coding tasks.
      3
      28 npm
      MIT
    • A
      license
      A
      quality
      D
      maintenance
      Optimizes token costs by intelligently delegating low-complexity tasks to local LLMs via LiteLLM, enabling cost-effective development workflows.
      3
      1
      MIT
    • A
      license
      A
      quality
      D
      maintenance
      Pre-execution cost estimation for LLM agent workflows, providing cost estimates before running tasks and improving accuracy over time through calibration.
      6
      30 PyPI
      2
      MIT