MCP Server

AccelsRouter ships a hosted MCP (Model Context Protocol) server — a single Streamable HTTP endpoint that exposes the gateway's catalog, pricing, account and chat abilities as tools your coding agent (Claude Code, Cursor, Codex, …) can call directly.

It's the same idea as OpenRouter's mcp.openrouter.ai: instead of remembering model ids and prices, your agent asks the gateway.

Endpoint & auth

Endpoint	https://router.accels.tech/mcp
Transport	Streamable HTTP (stateless)
Auth	Authorization: Bearer sk-... — your normal API key

The MCP endpoint runs behind the same authentication as the API. A disabled, expired, or exhausted key is rejected, and every tool acts as that key — you only ever see the models, prices, credit and request history that key is entitled to.

Setup

Claude Code

claude mcp add --transport http accelsrouter https://router.accels.tech/mcp \
  --header "Authorization: Bearer sk-xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx"

Cursor / generic JSON clients

Add to the client's MCP config (e.g. Cursor's mcp.json):

{
  "mcpServers": {
    "accelsrouter": {
      "url": "https://router.accels.tech/mcp",
      "headers": {
        "Authorization": "Bearer sk-xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx"
      }
    }
  }
}

Any MCP client that speaks Streamable HTTP with a custom header works the same way — point it at the endpoint and pass the Authorization header.

Tools

Seven of the eight tools are read-only and free. Only send-message consumes credit.

Tool	What it does	Key parameters	Cost
ping	Health check; returns ok and server time	—	Free
list-models	Models this key can call, with provider, supported endpoints and the effective price for this key	search, provider, endpoint, limit (≤500)	Free
get-model	Full details for one model: description, tags, provider, endpoints, billing type, effective price	model	Free
get-model-pricing	Effective price of one model for this key (USD per 1M tokens, or per call), including the group ratio	model	Free
get-credits	Remaining and used credit for this key (and the wallet that pays for it)	—	Free
get-generation	Cost, token counts and timing of one past request, by request id	request_id	Free
list-daily-model-rankings	Most-used models on the platform over the last N days, by token volume	days (≤30), limit (≤50)	Free
send-message	Send a chat message to a model and return the reply, token usage and cost	model, prompt or messages, system, max_tokens, temperature	Consumes credit

send-message is dispatched internally through the normal /v1/chat/completions path, so model allow-lists, routing, rate limits and billing apply exactly as they would for a direct API call.

Request ids

send-message returns a request_id (the same value as the X-Oneapi-Request-Id response header on direct API calls). Pass it to get-generation to read the final billed cost, token counts and latency once the call is recorded.

Typical workflow

A coding agent connected to the MCP server can, for example:

1. list-models — discover what models are available to the key
2. get-model-pricing — compare the effective price of two candidate models
3. get-credits — check there's enough balance
4. send-message — run a quick test prompt against the chosen model
5. get-generation — confirm the exact cost of that call

All of this happens without leaving the agent or hard-coding model ids and prices.

Notes

• Stateless — every request is a self-contained session; nothing is stored between calls.
• Same billing as the API — credit, quotas and rate limits are shared with your key's normal API usage; the MCP server is just another way in.
• Least privilege — issue a dedicated key (optionally quota-capped) for agents that only need catalog/pricing lookups. See API Keys.