DXT OpenRouter Router

A tiny MCP (stdio) server that returns a ready-to-send OpenRouter request body from a named preset and your current energy budget:

energy	Model slug becomes	Meaning
low	…:floor	cheapest provider serving that model
balanced	… (base slug)	the model's default provider
high	…:nitro	highest-throughput provider

Two things make this different from a normal cost router:

1. It routes on energy, not just cost. The input is a fact about you, not about the task. "Cheap" and "fast" are the same axis viewed from different energy levels.
2. Zero Data Retention is the default. Every body ships provider.data_collection: "deny", so a preset has to explicitly opt out of privacy rather than opt in to it.

It builds the request. It does not send it — so it never needs your API key in-process, and never sees a response.

---

Install

npm install -g dxt-openrouter-router

Or run it without installing:

npx dxt-openrouter-router

Register it with an MCP host

Claude Desktop (claude_desktop_config.json) or any MCP client:

{
  "mcpServers": {
    "openrouter-router": {
      "command": "npx",
      "args": ["-y", "dxt-openrouter-router"],
      "env": {
        "ROUTING_PRESETS_PATH": "C:\\path\\to\\routing.presets.json"
      }
    }
  }
}

ROUTING_PRESETS_PATH is optional — omit it and the bundled routing.presets.json is used. The file is re-read on every call, so you can edit presets without restarting the host.

The routeLLM tool

Argument	Required	Description
preset	✅	A key from routing.presets.json, e.g. research_long_context
energy		low | balanced | high (default balanced)
user_prompt		The user message to place in the body
system		Overrides the preset's own system prompt
overrides		Extra body fields merged in last, e.g. { "temperature": 0.2 }

Example

// call
{ "preset": "research_long_context", "energy": "low", "user_prompt": "Synthesize these sources..." }

// result
{
  "url": "https://openrouter.ai/api/v1/chat/completions",
  "method": "POST",
  "headers": {
    "Authorization": "Bearer $OPENROUTER_API_KEY",  // literal placeholder — never your real key
    "Content-Type": "application/json"
  },
  "api_key_configured": true,
  "body": {
    "temperature": 0.2,
    "model": "meta-llama/llama-3.1-70b-instruct:floor",
    "messages": [
      { "role": "system", "content": "You are a careful research synthesist. ..." },
      { "role": "user", "content": "Synthesize these sources..." }
    ],
    "provider": { "data_collection": "deny" }
  }
}

Presets

routing.presets.json is a plain map of name → preset:

{
  "presets": {
    "daily_driver": {
      "model": "google/gemini-3.6-flash",   // no :floor/:nitro here — energy adds that
      "description": "Default workhorse — unit tests, refactors, docs, CLI loops.",
      "system": "Optional default system prompt.",
      "provider": { "data_collection": "deny" },      // merged over the ZDR default
      "response_format": { "type": "json_object" },   // passed straight through
      "params": { "temperature": 0.4 }                // any other OpenRouter body field
    }
  }
}

Bundled presets mirror a simple decision tree:

Preset	Model	Reach for it when
quick_ping	x-ai/grok-4.5	formats, clarifications, vibe checks
daily_driver	google/gemini-3.6-flash	~80% of volume — tests, refactors, docs
logic_engine	openai/gpt-5.6-sol	math, symbolic logic, formal reasoning
high_reasoning	anthropic/claude-opus-4.8	architecture, nuanced review, agentic work
research_long_context	meta-llama/llama-3.1-70b-instruct	long-context synthesis across sources
structured_json	google/gemini-3.6-flash	extraction that must return parseable JSON

Slugs were verified against https://openrouter.ai/api/v1/models on 2026-07-22. OpenRouter slugs change as models ship — re-check before relying on them.

Secrets

🔒 This package contains no API key, and the tool output never includes one.

• OPENROUTER_API_KEY is read from the environment, and only ever reported as the boolean api_key_configured.
• The Authorization header is emitted as the literal string Bearer $OPENROUTER_API_KEY, so tool output is safe to paste into a chat log, an issue, or a commit.
• Copy .env.example → .env for local use. .env is gitignored.

Use as a library

The pure core is exported, so you can build bodies without MCP:

import { buildRequestBody, loadPresets } from "dxt-openrouter-router";

const presets = loadPresets("./routing.presets.json");
const body = buildRequestBody(presets, { preset: "daily_driver", energy: "low" });

Develop

npm install
npm run build     # tsc -> dist/
npm test          # node --test test/

License

MIT © Sasha Philius