// getting started

Connect your client

One endpoint, every client. The blocks below are generated from the same source the dashboard and the served setup prompt render from, so what you copy here is exactly what the platform would have told your agent to write.

Installing skills and plugins as well? The install matrix puts every client side by side, from timed installs.

01// the endpoint

One endpoint for everyone

mcp endpointtext
https://mcp.superagnt.com/mcp

The credential you present selects the workspace, not the URL. That means this string is identical for every customer: there is no per-tenant subdomain to look up and no path to personalize, and a config block copied from a teammate connects to your workspace as soon as you sign in.

Transport is Streamable HTTP MCP: JSON-RPC 2.0 over POST, with a stream that carries tools/list_changed when the served tool set changes. Clients that cannot speak remote HTTP MCP bridge through mcp-remote over stdio, which is the "Other" block below.

names

The server registers under the local alias superagnt. That alias is just your config file's name for the server and you can rename it freely. The tool prefixes (agnt_, data_, connection_, mcp_) are wire identifiers dispatched on by exact match and never change.
02// claude code

Claude Code

Two pieces. The MCP server goes in as a Claude connector: open the prefilled dialog on claude.ai and click Add. It reaches Claude Code signed in with your Claude account from the next session. The plugin carries the superagnt skills, the instructions for using the tools.

plugin · in a claude code sessionbash
/plugin marketplace add superagnt/plugins
/plugin install superagnt@superagnt

Claude Code signed in with an API key never loads connectors. There, add the server directly (connection only, no skills).

manualbash
claude mcp add --scope user --transport http superagnt https://mcp.superagnt.com/mcp
claude mcp login superagnt

--scope user is required

The CLI's default is project scope, which registers the server against the directory you happened to be in and hides it from every other one. The symptom is a connection that works in one repo and appears missing everywhere else.
03// per-client config

Every other client

Each block below is the whole change for that client. Where a config path is shown, create the file and its server map if they do not exist yet.

three silent schema traps

  • VS Code's key is servers, not mcpServers.
  • Gemini CLI's remote key is httpUrl, not url.
  • Windsurf uses serverUrl plus an explicit type.

All three accept a config copied from another client and then do nothing, with no error to read. Use the block for the client you are actually in.

Codex

oauth

run this from your terminal — oauth, no token needed

codexbash
codex mcp add superagnt --url https://mcp.superagnt.com/mcp
codex mcp login superagnt

Cursor

oauth

.cursor/mcp.json, then click Login under Settings → MCP

.cursor/mcp.jsonjson
{
  "mcpServers": {
    "superagnt": {
      "url": "https://mcp.superagnt.com/mcp"
    }
  }
}

VS Code

oauth

run this — the config key is "servers", not "mcpServers"

vs codebash
code --add-mcp '{"name":"superagnt","type":"http","url":"https://mcp.superagnt.com/mcp"}'

Gemini CLI

oauth

run this, then /mcp auth — the settings key is "httpUrl", not "url"

gemini clibash
gemini mcp add --transport http superagnt https://mcp.superagnt.com/mcp

Windsurf

oauth

~/.codeium/windsurf/mcp_config.json — "serverUrl" plus an explicit "type"

~/.codeium/windsurf/mcp_config.jsonjson
{
  "mcpServers": {
    "superagnt": {
      "type": "streamable-http",
      "serverUrl": "https://mcp.superagnt.com/mcp"
    }
  }
}

Zed

oauth

~/.config/zed/settings.json — accepts no headers, OAuth is the only path

~/.config/zed/settings.jsonjson
{
  "context_servers": {
    "superagnt": {
      "source": "url",
      "url": "https://mcp.superagnt.com/mcp"
    }
  }
}

Cline

oauth

run cline mcp, add it as a remote server, then choose Authorize OAuth

clinetext
https://mcp.superagnt.com/mcp

OpenClaw

oauth

run this from your terminal — oauth, no token needed

openclawbash
openclaw mcp add superagnt --url https://mcp.superagnt.com/mcp --transport streamable-http --auth oauth
openclaw mcp login superagnt

Hermes

bearer token

~/.hermes/config.yaml, then /reload-mcp

~/.hermes/config.yamltext
# ~/.hermes/config.yaml
mcp_servers:
  superagnt:
    url: "https://mcp.superagnt.com/mcp"
    headers:
      Authorization: "Bearer <YOUR_MCP_TOKEN>"

Other

bearer token

any client that speaks stdio via mcp-remote

otherjson
{
  "mcpServers": {
    "superagnt": {
      "command": "npx",
      "args": [
        "-y",
        "mcp-remote",
        "https://mcp.superagnt.com/mcp",
        "--header",
        "Authorization: Bearer <YOUR_MCP_TOKEN>"
      ]
    }
  }
}
04// credentials

OAuth or a bearer token

OAuth 2.1 is the lead path for almost every client, because the endpoint is its own authorization server. It publishes discovery metadata, accepts dynamic client registration, and runs PKCE S256, so an OAuth-capable client needs the URL and nothing else: it registers itself, you approve once in a browser, and the client holds its own credential. The consent screen doubles as signup, so a reader without an account still lands in a working workspace.

A workspace MCP bearer token from the dashboard is the fallback for clients that cannot do OAuth, and for people who would rather paste a header. Mint one at app.superagnt.com and drop it wherever the blocks above print <YOUR_MCP_TOKEN>.

never paste a real token here

Nothing on this page is a live credential. Every bearer slot is a placeholder, and a workspace MCP token is a secret with the same reach as your workspace. Keep it out of committed config files and out of screenshots.

The full credential contract, including what the REST API accepts and which scopes are refused where, is on Authentication.

05// gui-only

GUI-only clients

The Claude app, ChatGPT and Grok keep connectors in your account rather than in a file on disk. No agent can wire these for you, so the steps below are for a human: paste the endpoint into the connector form and approve the consent screen.

Claude app

oauth

add superagnt as a connector: open the prefilled dialog below and click Add

claude apptext
https://mcp.superagnt.com/mcp

ChatGPT

oauth

In ChatGPT on the web, go to Plugins, select the plus button, then Add custom MCP server. Name it superagnt, enter the server URL under Connection, and set Authentication to OAuth. Tick "I understand and want to continue", then select Create as a plugin. Install the plugin, then start a new chat, type @ and select superagnt. ChatGPT takes custom MCP servers on the web only, and OpenAI lists them for Business, Enterprise/Edu and Pro plans. Business and Enterprise/Edu workspaces get every tool, write actions included; an admin or owner adds the server there (on Enterprise/Edu, also a member they allow). Pro accounts get read and fetch actions only.

chatgpttext
https://mcp.superagnt.com/mcp

Grok

oauth

grok.com/connectors → New Connector → Custom. No token needed.

groktext
https://mcp.superagnt.com/mcp

account tier limits

Claude free accounts allow one custom connector. ChatGPT's plan rules are in its block above: on Pro the read tools answer and the write tools do not run.
06// verify

Verify it worked

The quickest independent check is a raw tools/list. It also doubles as the plain-HTTP integration path for anything that is not an MCP client at all. The Accept header carries both content types because Streamable HTTP lets the server answer either with a JSON body or with an SSE stream: a plain tools/list comes back as JSON, while a long-running tools/call that asked for progress is streamed back as text/event-stream.

tools/listbash
curl -sS https://mcp.superagnt.com/mcp \
  -H "Authorization: Bearer <YOUR_MCP_TOKEN>" \
  -H "Content-Type: application/json" \
  -H "Accept: application/json, text/event-stream" \
  -d '{"jsonrpc":"2.0","id":1,"method":"tools/list"}'

Inside a connected client, call agnt_onboarding instead. It confirms the connection and returns the platform brief in one move.

if the new tools are missing

Enabling a tool family changes the served surface immediately, but the calling client may keep showing the old list until it reconnects. Restart the client, or for an account-brokered connector refresh it in settings, and the new tools appear.
07// facets

Scoped entry points

The endpoint also answers on https://mcp.superagnt.com/mcp/<facet>. A server created now serves every curated data source whichever path you connect on: connecting through a facet for the first time provisions the workspace's default server with that facet's own tool families added, and stamps the slug so the install can be attributed to wherever it came from. A facet is an entry point, not a partition: once the server exists, your credential resolves the same server on any path. Servers created before the grouped tool list, which list one tool per operation, started scoped to their facet's families and keep that scope until you widen it, with no reconnect.

The facets match the MCP server listings: social-data, linkedin, x, instagram, tiktok, youtube, reddit, facebook, web-scraping, seo, lead-enrichment, email-finder and company-enrichment. A skill connects on the plain endpoint, and its setup guide switches on the tool families it needs.

An unregistered slug does not fail as a protocol error. It returns a structured 404 naming the valid facets, so a client that guessed can correct itself.