活动营销CRM MCP服务器

Nevent MCP

Talk to your live-events CRM (campaigns, analytics, paid ads, segments, short URLs) in Claude and ChatGPT.

Official MCP server for Nevent — the marketing platform for live event promoters. 63 tools across 10 groups (analytics, segmentation, campaigns, templates, deliverability, paid media, short URLs, media library, multi-tenant, and help/discovery) over OAuth 2.1 (hosted) or stdio (local).

📚 Full documentation: https://help.nevent.ai/en/nevent-ai/

Install

Hosted (recommended — no install needed)

Add https://mcp.nevent.ai/ as a remote MCP server in your client. Works with Claude.ai, ChatGPT, Cursor, Cline, Continue and any compatible MCP client.

→ Step-by-step with screenshots: Connect with Claude · Connect with ChatGPT

Local (stdio via npm)

npx mcp-nevent

Required env: NEVENT_JWT_TOKEN. Optional: NEVENT_OPERATION_MODE (READ_ONLY | STANDARD | FULL).

→ Full local setup: Local installation

Quick add

Claude.ai · ChatGPT (hosted, no setup)

Settings → Integrations / Connectors → Add custom MCP → URL https://mcp.nevent.ai/. Authorize with your Nevent admin account.

Claude Code

# Hosted (OAuth)
claude mcp add --transport http nevent https://mcp.nevent.ai/

# Or local (stdio)
claude mcp add nevent -- node /path/to/mcp-nevent/dist/index.js
export NEVENT_JWT_TOKEN=your_token

→ Full guide: Claude Code

Claude Desktop

Add to claude_desktop_config.json (macOS: ~/Library/Application Support/Claude/, Windows: %APPDATA%\Claude\):

{
  "mcpServers": {
    "nevent": {
      "command": "node",
      "args": ["/path/to/mcp-nevent/dist/index.js"],
      "env": {
        "NEVENT_JWT_TOKEN": "your_nevent_jwt_token"
      }
    }
  }
}

→ Full guide: Claude Desktop

Cursor · Cline · Continue · VS Code

→ Per-client setup: Cursor, Cline, Continue, VS Code

What you can ask

"Show me registrations for EventX broken down by ticket type"

"Which campaigns sent last month had an open rate above 30%?"

"Create a segment of attendees who bought a VIP ticket in 2025
 but haven't opened any email in the last 90 days"

"Schedule the 'Summer Gala' campaign for next Monday at 10:00 CET"

"Compare ROAS across our Meta ad campaigns for March"

"Create per-user tracking links for the Summer Gala campaign
 and show me which users clicked"

→ More patterns: Ready-to-use prompts · Use cases

Documentation

For developers

How it works

LLM (Claude / ChatGPT / Cursor / …)
        |
        | MCP (Streamable HTTP or stdio)
        v
  ┌─────────────────┐
  │  Nevent MCP     │  mcp.nevent.ai (hosted) or local
  └─────────────────┘
        |
        | REST + JWT
        v
  Nevent APIs (analytics, campaigns, segments, paid media, short URLs)

Hosted mode uses OAuth 2.1: the MCP server issues short-lived JWT tokens after validating your Nevent credentials. Each session is isolated — no shared service accounts.

Restricting OAuth clients (self-hosted HTTP)

Optional environment variables for the HTTP server:

See CHANGELOG.md for the exact behaviour.

Advanced: bearer-passthrough mode (trusted clients, service key required)

MCP_AUTH_MODE=bearer-passthrough is a second HTTP auth mode for trusted clients that already hold a per-user nev-api JWT and want to call MCP tools on that user's behalf — for example nev-helpbot, the Chatwoot support bot, answering as the promoter currently chatting.

MCP_AUTH_MODE=bearer-passthrough \
MCP_BEARER_PASSTHROUGH_SERVICE_KEY=$(openssl rand -hex 32) \
node dist/index.js --transport=http --port=3000

Every request must carry BOTH the client's service key and the calling user's own nev-api JWT:

X-Nevent-Service-Key: <pre-shared service key>
Authorization: Bearer <the promoter's nev-api JWT>

• Service key (mandatory). The instance may be reachable from outside, so a JWT alone is not enough. MCP_BEARER_PASSTHROUGH_SERVICE_KEY (at least 32 characters) is required: the server refuses to start without it. It is compared in constant time (crypto.timingSafeEqual over SHA-256 digests) and checked before the token is introspected, so outsiders cannot make this server call nev-api. A missing or wrong key gets a bare 401 (Unauthorized.) with no hint of which check failed. For rotation, set MCP_BEARER_PASSTHROUGH_SERVICE_KEY_PREVIOUS to the old key while the client moves to the new one, then remove it. Keep it in a secret store, never in the repository. Requests are also capped per IP (600/min, applied before body parsing) and per token (300/min). GET /health is unauthenticated for load balancers and returns only {status, version, commitSha}.

• Write policy (off by default). By default every session is READ-only (46 tools). MCP_BEARER_PASSTHROUGH_WRITES=human additionally exposes, to sessions of a person (principalType HUMAN verified in /auth/me), 12 write tools: nevent_create_segment, nevent_update_segment, nevent_create_template, nevent_update_template, nevent_clone_template, nevent_rename_template, nevent_send_test_template, nevent_create_campaign, nevent_schedule_campaign, nevent_create_short_url, nevent_update_short_url, nevent_create_bulk_user_short_urls. SERVICE sessions never get writes. upload_image, delete_image, reset_tenant, switch_tenant and list_tenants are never exposed, and an unclassified tool is denied. The policy is fixed per session at initialize. Writes are not safe to enable yet: the real barrier is the signed single-use write grant (NEV-1855), so the server refuses to start with human until that exists (override for non-production only: MCP_BEARER_PASSTHROUGH_WRITES_UNSAFE_NO_GRANT=true). Each tenant+user is limited to MCP_BEARER_PASSTHROUGH_WRITES_PER_HOUR writes per hour (default 30; over it, a write_rate_limit_exceeded tool error and nothing is executed; READ is never limited). Every tool call is logged with user_id, principal_type and the verified tenant (no email or token). Calls also need NEVENT_OPERATION_MODE=STANDARD for writes and FULL for nevent_schedule_campaign (classified DELETE because it sends for real); startup refuses human with READ_ONLY. The hourly limit is per process (in memory, each instance has its own budget, a restart resets it), counts only calls that actually execute (not invalid arguments, not calls the operation-mode guard rejects; a handler that fails does count) and accepts 1-1000. The no-grant override is rejected when NODE_ENV=production.

• No OAuth flow, no MCP_JWT_SECRET, no MONGODB_URI required. MONGODB_URI stays optional: set it to also enable the Mongo-backed read tools (campaigns/templates/deliverability), each still scoped by the caller's own JWT tenant.

• Every request is verified against nev-api before anything is exposed. On every request — initialize, tools/list, tools/call, not only at session start — this server calls GET {NEVENT_API_URL}/auth/me with the forwarded token. A 200 confirms the session is live and the request proceeds; 401, 403, a network error, or a timeout are all treated the same way — the MCP request is rejected with 401 and no tool, no tools/list, and no session is ever created or advanced from it. There is no permissive fallback: if nev-api cannot be reached to confirm the token, the request is refused, it is never let through anyway.

  • A successful check is cached briefly (~60 s) by a hash of the token so a single session does not call /auth/me on every message, but the cache never serves an entry past the token's own expiry.
  • The token's signature is still never checked locally by this server. nev-api signs session JWTs with a symmetric secret (HMAC256); shipping that secret here so it could be verified offline would mean compromising this process could forge a session for any Nevent user, not just read their data. Introspecting against nev-api avoids that trade-off entirely and is the only mechanism this mode uses to confirm a token is real, live, and unexpired. nev-data-api and nev-api additionally validate the token themselves on every downstream call they receive, exactly as they do for stdio mode's shared NEVENT_JWT_TOKEN.

• Human and service sessions. GET /auth/me also reports principalType (HUMAN or SERVICE), effectiveTenantId and serviceSession. Technical service sessions minted by POST /auth/service-token (unattended agents) are accepted like a human token only with TENANT scope (PLATFORM, i.e. global SUPERADMIN, is refused). The session is locked to effectiveTenantId, which comes only from this answer (never from the JWT payload; also correct for hierarchical admins navigating a child tenant). Fail-closed: an unknown principalType, a SERVICE response without effectiveTenantId, or any response with no tenant at all is refused. A SERVICE verification is cached for 10 s instead of 60 s so a disabled account or revoked session stops working almost immediately.

• Verification cache. Positive verdicts are cached (max 1000 entries, LRU, expired ones purged first); a token nev-api explicitly rejected (401/403) is remembered for ~5 s so a bad token cannot turn into one /auth/me call per request. When nev-api itself is unavailable (5xx, timeout, network) the request is still refused, but with 503 instead of 401, and that outcome is never cached.

• Read-only, tenant-locked. Only READ tools are exposed — the same set as an OAuth read-only client, including nevent_segment_execute — minus the tenant-switching tools (nevent_list_tenants / nevent_switch_tenant / nevent_reset_tenant): a bearer-passthrough caller always operates in the tenant its own JWT carries. Excluded and write/delete tools never appear in tools/list.

• Sessions are bound to their principal. Each Mcp-Session-Id remembers the user id, principal type and effective tenant it was opened for. Every later request (POST, GET, DELETE) must carry a token that nev-api confirms for that same principal and tenant, otherwise it is refused with 403 and nothing is dispatched — a valid token of someone else cannot drive a session it learned the id of. When the same principal presents a different valid token (a refreshed human token, or a service session re-minted after its 30 min expiry) the session's clients switch to the new token instead of keeping the one from initialize.

• Session isolation. Every MCP session gets its own client instances, built fresh from that session's own JWT — two sessions started with different tokens never share a client or cached data.

• Reachable from outside, authenticated twice. There is no login page and no client registration. Caller identity is the pre-shared service key (which client) plus the forwarded JWT (which end user), both required on every request. Still prefer a restrictive firewall/security group in front of it, leave MCP_ALLOWED_ORIGINS unset (the caller is a server, not a browser, so CORS is off and no Access-Control-Allow-Origin is emitted; set it only to a short explicit list if a browser client ever needs it), and never point the public OAuth hostname (mcp.nevent.ai) at a process started in this mode. Do not log or commit the service key.

Full design rationale: src/transports/http-bearer-passthrough.ts and src/auth/session-verifier.ts.

Privacy

The Nevent MCP server processes tenant data on Nevent's own infrastructure (AWS, eu-west-1). Data does not leave Nevent's infrastructure and is not used for model training.

For details on data collection, retention, third-party sharing, and contact information, see the full privacy policy: https://nevent.ai/en/privacy/

Security disclosure: see SECURITY.md.

Contributing

See CONTRIBUTING.md, CODE_OF_CONDUCT.md and SECURITY.md.

License

MIT © Nevent

Support

• Help Center: https://help.nevent.ai
• Issues: https://github.com/nevent-dev/mcp-nevent/issues
• Email: support@nevent.ai
• Product: https://nevent.ai/en