Why raw MCP configs become a security problem fast
Checking what staff paste into web chat (covered here) solves half the problem. Once teams connect AI assistants straight to business tools, three new risks show up:
- Passwords sitting on laptops: most MCP setup guides tell developers to paste long-lived GitHub, Jira or database tokens into a plain text file on every machine.
- All-or-nothing permissions: a token that lets an agent read tickets often lets it merge code, close tickets, message customers or delete records too. One bad prompt is all it takes.
- Nobody checks what comes back: even when the prompt is clean, a tool can pull customer records or secrets out of an internal system and hand them to a public AI model.
What works: put a checkpoint in the middle
Instead of pointing Claude Code, Cursor or Copilot straight at a business tool with a shared admin token, point it at your own Arx AI Agent Gateway address. The console includes 1-click presets for the services teams connect most often (such as GitHub, Jira and Confluence, Linear, Notion, Stripe, and custom internal or PSA gateways), and setup on the laptop is just a few lines of config:
{
"mcpServers": {
"github-prod": {
"type": "http",
"url": "https://app.arxcloud.io/mcp/your-tenant-id/github-prod",
"headers": {
"Authorization": "Bearer arx_agt_example_scoped_key"
}
}
}
}
- Keys stay off laptops: your real GitHub or Jira token is stored encrypted inside Arx Cloud. Each developer or workflow gets their own
arx_agt_...key that works for one connection and can be revoked in one click. If the real service rejects a request, the agent never sees the token or the login prompt. - You decide what each tool may do: pull in the list of tools a service offers, then mark each one as allowed, needs approval, or blocked. For example, allow
list_*andsearch_*, ask first forcreate_*ormerge_*, and blockdelete_*. New tools that appear later get a safe default based on what they sound like they do. Blocked tools are hidden from the assistant so it does not waste time trying them. - A human approves risky actions: when an agent tries something that needs approval, Arx Cloud pauses it and shows the request in your console. An admin clicks Approve Once, and the agent can retry that exact action one time within 30 minutes. The approval cannot be reused for a different action.
- Data is checked both ways: what the agent sends and what the tool returns are scanned for card numbers, personal data and API secrets. On Enterprise Hybrid and Sovereign, results are also matched against your real HIPAA, PCI & PII records. Sensitive values are masked or blocked before they reach the AI model, and every call lands in your audit log. Checks cover what your DLP rules define, so review those first.
- 1-click proxy pack blocks direct bypasses: enable the built-in Block Unbrokered Shadow MCP policy pack in one click so laptops cannot connect straight to public MCP endpoints around your gateway, alongside 1-click DLP packs for ChatGPT, Claude, Gemini, Copilot, and Cursor.
Scope and plan availability
To be upfront about what it covers today: the gateway works with public HTTPS MCP servers as well as private internal RFC1918 MCP servers inside an AWS VPC or on-prem datacenter (reached over an outbound-only Arx Private Access Publisher with zero inbound firewall ports). It cannot see MCP tools that run purely locally on a developer's own laptop (local stdio servers) unless you move them behind a central HTTP MCP service.
- Arx Business Team ($10/user/mo on 2-Year, 5+ seats, 30-day free trial): includes the AI Agent Gateway (up to 10 connections and 25 keys per connection), Arx Private Access Publishers for internal MCP servers, per-tool rules, human approvals, data scanning in both directions, and searchable console audit logs.
- Arx Enterprise Hybrid ($20/user/mo on 2-Year, 25+ seats, 30-day free trial): adds Exact Data Match (HIPAA, PCI & PII) on everything agents send and receive, real-time SIEM streaming, SAML single sign-on with contractor session watermarking, and dedicated VPC relays.
- Arx Enterprise Sovereign ($30/user/mo on 2-Year, 75+ seats, 30-day free trial): adds Dedicated / Sovereign VPC isolation, unlimited vaulted MCP connectors, custom HIPAA BAA / DPA redlines, security questionnaires, and a Dedicated Solutions Engineer.