MCP (Model Context Protocol) server for ArcRouter — use intelligent LLM routing directly from Claude Code, Cursor, Cline, and other MCP-compatible AI tools.
Add ArcRouter to your MCP client in one command:
claude mcp add arcrouter --transport http https://api.arcrouter.com/mcpOr add to your MCP config manually:
{
"mcpServers": {
"arcrouter": {
"url": "https://api.arcrouter.com/mcp",
"transport": "http"
}
}
}The server exposes 6 tools:
Route a prompt to the best model.
| Parameter | Type | Description |
|---|---|---|
prompt |
string | The message to send |
budget |
string | free | economy | auto | premium |
mode |
string | default | council |
session_id |
string | Session ID for model pinning |
Returns: Response text, model used, topic detected, cost estimate.
Example:
Use arcrouter to explain how transformers work, with budget=premium
List models with benchmark scores and pricing.
| Parameter | Type | Description |
|---|---|---|
topic |
string | Filter by topic (e.g. code, math, science) |
budget |
string | Filter by budget tier |
limit |
number | Max results (default 10) |
Returns: Model list with quality scores, cost per 1M tokens, context length.
Check ArcRouter system status.
Returns: Health status for database, routing, and all 5 direct providers (OpenAI, Anthropic, Google, DeepSeek, xAI).
Run a prompt through council mode — multiple models answer in parallel and a chairman synthesizes verdicts. Higher-confidence answers for high-stakes questions; slower than arcrouter_chat (5-15s).
| Parameter | Type | Description |
|---|---|---|
prompt |
string | The question to verify |
budget |
string | Council member quality (free/economy/auto/premium) |
Returns: Synthesized answer + per-model votes + agreement percentage.
Get usage stats for a workflow session — total spend, models used, latency, tier distribution.
| Parameter | Type | Description |
|---|---|---|
session_id |
string | Workflow session ID (from workflow_budget.session_id) |
Returns: Spend tracker + budget remaining + tier distribution. No auth required — session_id is the secret.
Get cumulative usage stats for an ArcRouter API key.
| Parameter | Type | Description |
|---|---|---|
api_key |
string | Your ArcRouter API key (sk_...) |
days |
number | Number of days to query (1-90, default 30) |
Returns: Daily request counts + total cost summary.
The MCP server supports four authentication tiers:
- Free tier: No auth required. Uses free models only (rate-limited 30 req/hr).
- API key: Pass
Authorization: Bearer sk_...for paid tier (1000 req/hr). - x402 (USDC on Base): Pass
X-PAYMENTheader for per-request crypto payment. - MPP (USDC.e on Tempo): Pass
Authorization: Payment <credential>header. Sub-cent fees, 500ms finality. See mpp.dev.
ArcRouter routes LLM requests to the best model based on:
- Topic detection — code, math, science, writing, reasoning (24 categories)
- Complexity scoring — SIMPLE/MEDIUM/COMPLEX/REASONING tiers
- Benchmark scores — 345+ models scored on LiveBench, LiveCodeBench, HuggingFace
- Cost optimization — budget tiers from free to premium
- API: https://api.arcrouter.com
- Docs: https://arcrouter.com/docs
- SDK:
npm install arcrouter - Classifier: https://github.com/ArcRouterAI/arcrouter