SSentinelDocumentation
Open Sentinel
Tools & MCP

Build

Tools & MCP

Give agents controlled access to external capabilities.

Tools workspace

Sentinel Tools page with an active MCP connector and read-only custom tool
Connectors and custom tools share one workspace but keep separate lifecycles and contracts.
  1. 1Creation actionsAdd an MCP connector or define a custom tool.
  2. 2Connector healthServer URL, status, identity, refresh/connect, remove, and discovered-tool count.
  3. 3Custom tool contractName, description, category, status, input parameters, output schema, and usage metadata.
  4. 4Operational statusUse active/idle/error state before assigning the tool to an agent.

Tool types

Custom tool

A named action with an input contract and implementation you control.

MCP connector

A remote MCP server whose tools Sentinel can discover and assign.

Authenticated connector

An MCP connection using automatic OAuth, an API key header, or intentionally no authentication.

Tool test

A recorded test run used to verify inputs, outputs, latency, and failure behavior.

Connector lifecycle

  1. 1

    Add the server

    Enter a descriptive name and HTTPS MCP server URL.
  2. 2

    Authenticate

    Use automatic OAuth when supported, an API key only in the expected header, or no authentication only for a public server.
  3. 3

    Discover

    Sentinel performs the MCP handshake and lists the server's capabilities.
  4. 4

    Refresh or reconnect

    Use the connector control when tools change or authentication expires.
  5. 5

    Assign narrowly

    Attach only the required connector to the target agent, then test in Playground.

Custom tool contract

  • Use a verb-oriented name and describe exactly when the agent should call it.
  • Define typed parameters and mark only truly mandatory inputs as required.
  • Document whether the operation is reversible and any approval boundary.
  • Define an output schema so downstream reasoning receives predictable data.
  • Monitor total invocations, average latency, success rate, and recent usage.

Tool analytics and test runs

Custom tools and connectors are backed by analytics, version, and test-run endpoints (/tools/<tool-id>/analytics, /tools/<tool-id>/versions, /tools/<tool-id>/test-runs). The Tools page does not surface these panels yet, so call the endpoints from your own tooling. Review invocation count, average latency, success rate, and recent usage before assigning a tool more widely, and run a recorded test call to verify inputs, outputs, latency, and failure behavior without touching a live agent.

  • Check invocation volume before giving a tool to more agents — a spike can signal an over-broad prompt.
  • Use test runs to confirm timeouts and error messages are understandable.
  • Treat a tool&apos;s analytics and version history as part of the audit trail.

Permissions