Fresh 2026

Agent Tool Call Flow

How a tool call from an AI agent flows through Agent Handler: MCP, per-user auth, security scanning, Connector dispatch, and logging.

Agent Handler is an MCP server that sits between your agent and hundreds of third-party Connectors. It adds the layer a bare MCP server lacks: per end-user authentication, credential storage, scoping, data-loss prevention, and an audit log.

flowchart TD
    A[Agent: tools/call via MCP] --> B[Resolve bearer token]
    B --> C[Org + Tool Pack + Registered User]
    C --> D[Attach that user's stored credentials]
    D --> E[Security: scope + DLP scan]
    E --> F{Allowed?}
    F -- No --> G[Block + log denial]
    F -- Yes --> H[Dispatch to Connector]
    H --> I[Third-party tool executes]
    I --> J[Normalize response]
    J --> K[Audit log: who, what, result]
    K --> L[Return tool result to agent]

Stages

1. Initialize the session. The agent connects to the Agent Handler MCP endpoint with a bearer token that resolves to an organization, a Tool Pack, and a Registered User.
2. List tools. The agent calls tools/list to see the tools in its Tool Pack.
3. Call a tool. On tools/call, Agent Handler attaches the right user's credentials, no secrets ever reach the agent.
4. Security pass. The call is checked against your scoping and data-loss-prevention rules. See Secure.
5. Connector dispatch. The request runs against the third-party Connector. See Connectors.
6. Observe. Every call is recorded with who ran it, what ran, and the result. See Observe.

What Agent Handler adds on top of MCP

Capability	Bare MCP	Agent Handler
Tool listing + calling	Yes	Yes
Per end-user auth	No	Yes
Credential storage	No	Yes
Scoping + DLP	No	Yes
Audit log	No	Yes