MCP Servers

What it does

MCP Servers is where you register Model Context Protocol (MCP) servers — the part of Arbitex’s MCP Security channel module that lets you connect external tool servers to agents. For each server you set a name, endpoint, transport, and an optional auth token, then author which of its tools are explicitly allowed or denied.

Configure in the UI

Models & Routing → MCP ServersCitadel synthetic data on a non-production Arbitex demo instance. Never client data.

1. Open Models & Routing → MCP Servers and select Add server.

2. Fill in Name, URL, and Transport (stdio, HTTP, or SSE). Auth token is optional and masked — it is write-only; editing a server never re-displays the stored token, and leaving the field blank keeps the existing one.

3. Under Tool policy, add a row per tool the server exposes and mark it Allow or Deny. Tools with no row are neither explicitly allowed nor denied.

4. Toggle Enabled off to keep a server’s registration without exposing it — a disabled server is reported unavailable to the gateway. Select Add server to save.

5. The table’s Health column shows the live status from the gateway’s network probe (Healthy / Degraded / Unreachable / Unknown) when available, or falls back to a Status column reflecting the authored Enabled/Disabled state if the live probe isn’t wired for your deployment.

6. Click a row to expand its Tool manifest and review allow/deny badges per tool at a glance; use the row menu for Edit or Delete.

Policy eval-test is a preview, not a live check

The Policy eval-test panel lets you paste a payload, pick a direction (Input (request) / Output (response)), and name a tool, then select Run eval to preview the decision the gateway would return. This dry-run runs against the platform gateway, not the admin control plane — wiring it up is a tracked follow-on, so today it always reports that the decision path is not connected rather than fabricating a pass/fail result.

API equivalent

curl -X POST "https://admin.arbitex.ai/api/v1/admin/models-routing/mcp-servers/list" \
  -H "Authorization: Bearer $ADMIN_TOKEN" \
  -H "Content-Type: application/json" \
  -d '{"du_id": null}'

curl -X POST "https://admin.arbitex.ai/api/v1/admin/models-routing/mcp-servers/set" \
  -H "Authorization: Bearer $ADMIN_TOKEN" \
  -H "Content-Type: application/json" \
  -d '{"du_id": null, "objects": [{"object_key": "internal-docs", "name": "internal-docs", "url": "https://mcp.internal.example.com", "transport": "http", "enabled": true, "tool_policy": [{"tool": "search", "policy": "allow"}]}]}'

curl -X POST "https://admin.arbitex.ai/api/v1/admin/models-routing/mcp-servers/health" \
  -H "Authorization: Bearer $ADMIN_TOKEN" \
  -H "Content-Type: application/json" \
  -d '{"du_id": null}'

set authors the complete server set (including deletes) as one revision; there is no partial-update verb.

Plan & availability

MCP Servers is available on every plan as part of the MCP Security channel module.

Deployment notes

On a Hybrid Outpost, MCP servers are registered per deployment unit, and the auth token you supply is delivered to the gateway with the registration — it is never re-displayed by the console after you save it. Confirm an outpost’s connectivity on Fleet & Outpost before registering a server whose endpoint is only reachable from that outpost’s network.

Related

• Agent Tool Policies — a separate, not-yet-enforced tool-policy surface for agents generally.
• Retrieval Sources (RAG) — another typed connection surface for grounding content.