Self-hosted MCP gateway

Every connector.
One endpoint.
Zero glue code.

Connect LINE, Gmail, GitHub and the rest of your stack once. ConnectorX aggregates their tools behind a single secure MCP URL for Claude, Cursor, or any AI agent — with per-connection permissions, write gates and metering built in.

LINEGmailGitHubZimbraTradingViewAcctoolXEkkasarn
terminal — connect Claude
$ claude mcp add --transport http connectorx https://mcp.connectorx.ai/mcp
 connectorx wired — sign in once, approve your connections
 200+ tools ready: line_*  gmail_*  github_*  zimbra_*  …
> reply to the LINE customer asking about opening hours
 line_replyMessage · Shop OA · write gate open
 replied in 1.4s · logged to Activity
Use cases

What will your agent actually do?

Shop & customers

Answer the FAQ, take the booking, turn a chat order into a bill, nudge the lead that went quiet a week ago.

across
Personal secretary

Voice-note a thought on the go, snap a receipt, ask “ยอดเดือนนี้เท่าไหร่” — and get a digest every morning at 7.

across
Inbox zero

Triage overnight mail, draft replies in your voice, label and archive the noise before you've had coffee.

across
Thai back office

Issue the quotation, send the invoice, generate the official memo — number formats and forms done right.

across
Engineering

Open the issue, paste the repro, comment the fix — and ping the shop-floor chat when CI goes red.

across
Markets

Read the live chart, run the screener, capture the setup for tomorrow's review — on your own desktop session.

across
Capabilities

A permission layer, not just a proxy.

01

Write gates

Every connection starts read-only. Sends, edits and broadcasts refuse to run until a human flips Allow writes — and the gate applies on the very next call.

  • Per connection, not per account
  • Revocable any time — no reconnect needed
  • Reads always keep working
write gate — LINE · Shop OA
$ line_broadcastMessage {…}
✗ write_not_allowed — connection is read-only
Allow writes · ON
$ line_broadcastMessage {…}
sent to 1,204 followers
02

Scoped sessions & OAuth 2.1

Mint expiring MCP surfaces narrowed to specific connectors, tools or accounts — and let remote clients authorize through branded OAuth consent instead of pasted tokens.

  • Sessions expire; scope only narrows
  • Full OAuth 2.1 authorization server built in
  • Device tokens for CLIs and legacy clients
scoped session
POST /api/sessions {"scope":{"connectors":["line","gmail"]}}
201 mcps_9f3…e21 · expires in 24h
client tool list: 26 tools (of 200+)
remote client? → branded consent at connectorx.ai
03

Metering & credits

1,000 free tool calls a month on every plan, then flat per-call credits with an append-only ledger. Reserved before the call, refunded when it fails.

  • Failed calls are free — auto-refunded
  • Prepaid credits, no surprise bills
  • PromptPay, cards and Thai-baht rails
usage — this month
free calls 812/1,000
credit balance 5,000
✗ gmail_send_message timed out refunded ✓
ledger: append-only · every movement audited
04

LINE inbound inbox

Beyond push: customer messages, voice notes and slip photos arrive through a signed webhook inbox your agent can poll, download and answer.

  • Signature-verified webhook you own
  • Voice notes & photos downloadable to the agent
  • Single-use reply tokens handled for you
inbox — LINE · Shop OA
POST /webhooks/line · voice note 0:07
getRecentEvents 1 new since #4821
line_replyMessage "เปิด 9:00–18:00 ทุกวันครับ"
replied · reply token used once
How it works

Three steps, then it's just tools.

01

Connect once

Add each integration in the dashboard — credentials are AES-256-GCM encrypted and never leave your host. Multiple accounts per connector, each with its own permissions.

02

Authorize your agent

Add the endpoint as a custom connector in Claude for branded OAuth consent, or mint a scoped device token for CLIs and legacy clients.

03

Ship

Your agent sees every approved tool, namespaced per connector — line_replyMessage, gmail_send_message — with health, activity and usage tracked per call.

Self-hosted · Security

Yours, end to end.

ConnectorX runs as one Node process on your own hardware behind a Cloudflare tunnel. Credentials never leave your host, nothing is brokered through a third party, and the permission layer fails closed.

Health is reactive — a failure flags the connection, the next success heals it, and one-click Test probes stored credentials on demand.

connection: active
guarantees
AES-256-GCM credential encryption at rest
write tools refuse until a human opens the gate
tool args & results never stored — metadata-only activity log
scoped sessions expire · OAuth grants revocable
your database, your keys, your hardware

Point your agent at one URL.

Open dashboard
MCP endpoint  https://mcp.connectorx.ai/mcp