Get started

Quickstart: MCP

Connect the MCP server: the discover, inspect, run loop as native agent tools.

RouterGrowth speaks Model Context Protocol: connect once and your agent gets the discover, inspect, run loop as native tools, with the spending rules delivered in the handshake.

https://api.routergrowth.com/mcp

Streamable HTTP transport, stateless: no session to manage, works from any MCP client that can send an Authorization header.

Authentication

The server authenticates with your normal API key in the Authorization header, exactly like REST. Mint a key on the dashboard. discover works without a key; every other tool returns a clear message telling the agent how to get one, so a keyless connection still lets you browse the catalog.

Connect a client

Claude Code

claude mcp add --transport http routergrowth https://api.routergrowth.com/mcp \
  --header "Authorization: Bearer rg_live_..."

Cursor, VS Code, Windsurf, Cline and any client using an mcpServers config:

{
  "mcpServers": {
    "routergrowth": {
      "type": "http",
      "url": "https://api.routergrowth.com/mcp",
      "headers": {"Authorization": "Bearer rg_live_..."}
    }
  }
}

Codex

codex mcp add routergrowth --url https://api.routergrowth.com/mcp \
  --header "Authorization: Bearer rg_live_..."

Claude.ai custom connectors require OAuth sign-in, which lands next; until then use a client that sends headers, or the skill, which covers every platform today.

The tools

Tool	Auth	What it does
`discover`	none	search the catalog by job description; free, never runs a paid call
`inspect`	key	schema, providers, exact price and billing conditions for one capability
`run`	key	execute a capability; billable, with `max_cost`, routing and `idempotency_key`
`runs`	key	run history with attempts, charges and routing reasons
`get_run`	key	one run by id
`balance`	key	wallet balance, reserved and available

The tools call the same code as /v1: identical routing, billing, waterfall and error codes, with errors returned as tool results the model can read and react to.

Spending rules

The server's initialize response instructs the agent: inspect and surface the price before any billable run, set max_cost on every call, ask the user before a batch that would exceed about $1, and never present sandbox (rg_test_) mock data as real. The same rules ship in SKILL.md.

Quickstart: agents: the skill, for platforms without MCP.
Run reference: what the run tool does under the hood.

Previous
Quickstart: agents Next
API overview

Reading this as an agent? This page as markdown: /docs/quickstart-mcp.md · every page: /docs/llms.txt