# Connect RouterGrowth to your agent

Connect once, then verify with free calls. Setup does not require credit or execute a provider tool.

- **MCP endpoint:** `https://api.routergrowth.com/mcp`
- **Transport:** Streamable HTTP
- **Authentication:** OAuth where supported; an API key for headless clients.
- **Account and keys:** [RouterGrowth dashboard](https://www.routergrowth.com/dashboard)

This page is a setup guide. Enter the MCP endpoint above in connector settings, not this page's URL.

## 1. Use the intended client

Follow the app and environment the user selected. If none was named, use the current app when it is clear; ask only when the target is ambiguous. Configure just that client.

Check for an existing RouterGrowth connection first, including one supplied by a plugin. Reuse a connection to the same endpoint regardless of its name. Preserve other servers and settings, and do not replace a different connection under the same name without asking. Keep the client's approval controls.

## 2. Connect and authorize

Choose one recipe below. Let the user complete browser sign-in and workspace consent. A saved configuration alone does not prove authorization succeeded.

| Client | Setup |
|---|---|
| Codex | Run `codex mcp add routergrowth --url https://api.routergrowth.com/mcp`. Complete the browser flow if it starts; otherwise run `codex mcp login routergrowth`. For an existing connection, use its name. [Details](https://www.routergrowth.com/docs/quickstart-mcp#codex). |
| Claude Code terminal | Run `claude mcp add --transport http routergrowth https://api.routergrowth.com/mcp`, then open `/mcp` in Claude Code and authenticate. [Scope and key options](https://www.routergrowth.com/docs/quickstart-mcp#claude-code). |
| Claude Desktop, Code tab | For a local project, use its Connectors UI or merge an HTTP server with the endpoint above into the project's `.mcp.json`, preserving existing entries. Complete OAuth in that session. [Details](https://www.routergrowth.com/docs/quickstart-mcp#claude-code). |
| Claude web, Desktop Chat/Cowork, or mobile | Add RouterGrowth through **Customize → Connectors → Add custom connector**, using the MCP endpoint. Connect, approve the workspace, then enable it in the chat. [Details](https://www.routergrowth.com/docs/quickstart-mcp#claudeai-claude-desktop-and-claude-mobile). |
| ChatGPT | Add a custom MCP connection with the endpoint above and OAuth, then select it in the conversation. [Current setup and workspace requirements](https://www.routergrowth.com/docs/quickstart-mcp#chatgpt). |
| Cursor, Hermes, other IDEs or clients | Use the [client-specific recipe](https://www.routergrowth.com/docs/quickstart-mcp#pick-your-setup); prefer OAuth where supported. |
| Headless runtime or CLI | Use the [API-key setup](https://www.routergrowth.com/docs/quickstart-agents). Keep the key in the runtime's secret store or private environment. |

For a remote or cloud session, configure the environment that actually runs the agent. Local settings may not reach it. If the client has no custom connector support or the workspace blocks it, report that limitation.

Never ask for a key in chat, print credentials, or save them in a repository. If an API key is needed, the user creates it privately in the dashboard. An `rg_test_` key returns simulated data; it is not a live-provider connection.

## 3. Verify without spending

Run these through the connection in the intended agent session:

1. Call `balance` with `{}` to confirm workspace authentication. A zero balance is fine.
2. Call `discover` with `{"query":"web search","kind":"capability","limit":3}`.
3. Take a capability slug returned by that call and call `inspect` with `{"capability":"<returned slug>"}`. Do not invent an identifier.

All three calls are free. For an existing CLI connection, use `routergrowth balance`, `routergrowth discover -q "web search" --kind capability --limit 3`, then `routergrowth inspect -c <returned-slug>`.

Report **configuration saved**, **account authorized**, and **connection verified** only as each succeeds. Verification requires successful tool results, not just a successful command or HTTP response. Include the inspected capability, its availability, required inputs and price; distinguish any reservation from the final charge. Label test-key results as simulated.

If tools are not available yet, reload or reconnect the intended session and retry these free checks. If authorization, network access or a tool call fails, name the remaining step instead of claiming the connection works. Public catalog discovery alone does not verify workspace access.

Do not call `run`, `batch_run` or `watch`, fund a wallet, or change spending controls during setup. Once verified, setup is complete. If the user also requested a job, continue under the [RouterGrowth skill](https://www.routergrowth.com/SKILL.md) and that job's authorized scope.