@nsupp/mcp

Model Context Protocol server for nsupp — lets an MCP client read conversations, draft and send replies, look up contacts and post to team chat. Dependency-free, Node 18+, stdio transport.

Configure

Add it to your MCP client. The secret goes in env, never on the command line:

{
  "mcpServers": {
    "nsupp": {
      "command": "npx",
      "args": ["-y", "@nsupp/mcp"],
      "env": {
        "NSUPP_TOKEN_ID": "nsupp_pk_…",
        "NSUPP_TOKEN_SECRET": "…",
        "NSUPP_WEBSITE_ID": "wk_…"
      }
    }
  }
}

NSUPP_BASE_URL points it at a self-hosted install; NSUPP_TIER selects plugin (default) or website tokens.

Tools

| Tool | Reads / writes | | --- | --- | | get_website, list_operators | read | | list_conversations, get_conversation, list_messages | read | | send_message | writes to the customer | | add_internal_note, set_conversation_state | write, internal | | list_people, get_person | read (personal data) | | list_team_chat, post_team_chat | general team channel | | list_helpdesk_articles | read | | list_bot_blocks, validate_bot_graph, list_bot_scenarios | read (bot authoring) | | create_bot_scenario, update_bot_scenario | write — draft only, no visitor sees it | | deploy_bot_scenario | publishes the bot to real visitors |

Every tool declares the scope it needs in its description, so a 403 tells the model exactly which permission to ask for. Write tools carry MCP annotations — send_message is marked readOnlyHint: false, destructiveHint: true so your client can require a human confirmation before a message reaches a customer. deploy_bot_scenario carries the same annotation for the same reason: publishing a scenario makes it, from that moment, the script every visitor on the selected channels talks to. Authoring is deliberately a four-step pipeline — list_bot_blocks (learn the contract) → validate_bot_graph (check without saving) → create_bot_scenario (draft) → deploy_bot_scenario (publish). Skipping the first step is the most common cause of a bot that silently never runs, because a guessed block name is not a valid block.

Boundaries worth knowing

• Team chat is the general channel only. Private groups and direct messages are not exposed — their membership belongs to a person, and an API key has no person behind it.
• The server decides, not this process. Plan tier, scopes, granted consent and tenant isolation are enforced in nsupp's own gate. A model calling a tool has exactly the authority the token has, never more; a 402 plan_required reaches the model verbatim instead of being smoothed over.
• stdout carries protocol only. Diagnostics go to stderr.

Apache-2.0