Connect a client - GetSales

The server speaks streamable HTTP at https://mcp2.getsales.io/mcp and authenticates with OAuth. Your client discovers the flow on its own: the first call comes back with a 401 and a WWW-Authenticate challenge pointing at the server’s metadata, the client registers itself (dynamic registration, nothing to pre-register), and you approve the GetSales consent screen in your browser: sign in, pick the workspaces the client may act in, click Allow. Your password never reaches the client.

Add to your agent

One action per client, then approve the consent screen that opens. That is the whole setup.

Claude

Open Settings, Connectors, then “Add custom connector”. Name it GetSales, paste https://mcp2.getsales.io/mcp, Add, then Connect and approve access. Works for claude.ai and Claude Desktop.

Cursor

One-click install (the button opens Cursor). Press Connect on the server and sign in on the consent page that follows.

VS Code

One-click install (the button opens VS Code). Sign in on the consent page that follows.

ChatGPT

Settings, Connectors, Create: URL https://mcp2.getsales.io/mcp, authentication OAuth. It registers itself and opens the consent screen like the others. On Business and Enterprise workspaces an admin may need to allow custom connectors first.

Claude Code

claude mcp add --transport http getsales https://mcp2.getsales.io/mcp

Then run /mcp in any session and authenticate when prompted. claude mcp list shows getsales as connected once consent is done. One domain mount instead of the facade puts every tool of that domain in the tool list, with no discovery step:

claude mcp add --transport http gs-tables https://mcp2.getsales.io/mcp/playtables/tables

Codex

codex mcp add getsales --url https://mcp2.getsales.io/mcp
codex mcp login getsales

The first line registers the server, the second runs the OAuth flow and opens the consent page.

Claude Desktop and Cursor by file

A client that configures servers by file takes the bare URL:

{
  "mcpServers": {
    "getsales": { "url": "https://mcp2.getsales.io/mcp" }
  }
}

Any other MCP client

Point it at https://mcp2.getsales.io/mcp over the streamable HTTP transport and let it run the OAuth flow. The discovery documents are standard: the protected-resource metadata at https://mcp2.getsales.io/.well-known/oauth-protected-resource (RFC 9728) names the authorization server, whose metadata is at https://beta.getsales.io/.well-known/oauth-authorization-server/id/v1 (RFC 8414; the issuer https://beta.getsales.io/id/v1 has a path, so its document lives at that root-inserted form).

Clients that only spawn a command

Some clients cannot take a URL at all, only a command to run over stdio. The open mcp-remote bridge runs the same OAuth flow, opens the consent page on first use and caches the tokens under ~/.mcp-auth:

{
  "mcpServers": {
    "getsales": {
      "command": "npx",
      "args": ["-y", "mcp-remote", "https://mcp2.getsales.io/mcp"]
    }
  }
}

Headless: authenticate with an API key

The same key that authenticates the REST API works on this server as a plain bearer token. No browser, no consent screen, so it fits CI, servers and any agent runtime where the OAuth round trip is not an option. Mint a key in the app (Settings, API keys) and attach it as a header:

claude mcp add --transport http getsales https://mcp2.getsales.io/mcp --header "Authorization: Bearer gs_live_..."

For command-spawning clients, the bridge takes the same header:

{
  "mcpServers": {
    "getsales": {
      "command": "npx",
      "args": ["-y", "mcp-remote", "https://mcp2.getsales.io/mcp", "--header", "Authorization: Bearer gs_live_..."]
    }
  }
}

Pick one mode per connection. A client that attaches an Authorization header and also runs the OAuth flow overwrites the OAuth token with the static header on every request, and the connection fails in a way that looks like a server problem. With a key attached the OAuth flow never starts; without one it always does. The key is scoped to its workspace and its permission set, revocable in the app, and treated like a password (see Authentication). Interactive clients such as claude.ai connectors have no header field and always use OAuth.

Check it from a terminal

A key is the simplest bearer for a terminal. Every call is one JSON-RPC message; the Accept header has to name both media types.

# find a tool by what it does, with its whole definition
curl -sS -X POST https://mcp2.getsales.io/mcp \
  -H "Authorization: Bearer gs_live_..." \
  -H 'Content-Type: application/json' -H 'Accept: application/json, text/event-stream' \
  -d '{"jsonrpc":"2.0","id":1,"method":"tools/call","params":{"name":"search_tools","arguments":{"query":"pause a workflow","detail":"full"}}}'

# one domain mount, the first page of tables
curl -sS -X POST https://mcp2.getsales.io/mcp/playtables/tables \
  -H "Authorization: Bearer gs_live_..." \
  -H 'Content-Type: application/json' -H 'Accept: application/json, text/event-stream' \
  -d '{"jsonrpc":"2.0","id":2,"method":"tools/call","params":{"name":"search_playtables","arguments":{"page_size":5}}}'

To walk the OAuth flow by hand instead, run npx @modelcontextprotocol/inspector, choose the streamable HTTP transport, enter https://mcp2.getsales.io/mcp, connect, and complete the consent screen. The tool list appears once you are authorized.

What you get

On the facade, four meta-tools rather than hundreds of individual ones: search_tools to find tools by what they do, describe_tool to read one definition, call_tool to run any tool by name, and list_toolsets to see the map. Have the full definition before the first call: search_tools with detail: "full" when the pick is clear, describe_tool otherwise, then call_tool. On a domain mount the tools are listed directly. Dangerous tools still run their preview and commit step, so an agent cannot delete a table or cancel a run without showing you what it is about to do.

First prompt to try

“List my tables and tell me which data sources are still importing. Then find the three newest conversations with a reply nobody answered yet and draft an answer for each, without sending.”

The agent chains the tools on its own; the drafts wait for your word.

​Add to your agent

Claude

Cursor

VS Code

ChatGPT

​Claude Code

​Codex

​Claude Desktop and Cursor by file

​Any other MCP client

​Clients that only spawn a command

​Headless: authenticate with an API key

​Check it from a terminal

​What you get

​First prompt to try