Taskriz Adds Human-in-the-Loop API for AI Agents

Overview

Taskriz has launched a new Human-in-the-Loop API designed to bridge the gap between autonomous AI agents and human decision-makers. The update allows developers to integrate structured human responses directly into their agent workflows, enabling systems to pause for clarification or approval before proceeding. This development marks a significant step toward more reliable and accountable AI operations, particularly in enterprise environments where human oversight is mandatory.

The core innovation lies in Taskriz’s ability to transform complex natural language requests into intuitive user interfaces that humans can interact with in seconds. By offering both plugin integrations for popular tools like Claude Code and ChatGPT, as well as a standalone API for custom workflows, the platform aims to make human feedback loops accessible regardless of a developer’s technical stack.

"Developers can now give their agent a way to ask people and receive structured answers back," Taskriz stated. "This transforms AI agents from static scripts into dynamic collaborators capable of seeking guidance when needed."

Seamless Integration via Plugins

For developers already embedded in the Claude Code and ChatGPT ecosystems, Taskriz offers immediate integration through dedicated plugins. These tools are designed to lower the barrier to entry, allowing agents to create human-assistance requests without writing custom code.

Installation is straightforward for both platforms. Claude Code users can add the plugin via the marketplace command, while ChatGPT users utilize their respective plugin installation commands. Both plugins connect to the Taskriz MCP server at https://www.taskriz.com/mcp and require an OAuth sign-in during the agent’s first request.

Once authenticated, agents can design request interfaces and send them to designated humans. The process is nearly invisible to the end user; the agent simply triggers a request, and the human recipient receives a structured prompt through the Taskriz interface. Agents created through these plugins are automatically linked to the user’s Taskriz account and associated subscription plan.

This approach is ideal for rapid prototyping and for teams that want to add human oversight to existing conversational AI workflows without disrupting their current tooling. It effectively turns any Claude or ChatGPT session into a hybrid system that can pause and consult human experts.

The Taskriz API for Custom Workflows

For organizations requiring deeper customization or those building agents outside the major consumer platforms, Taskriz provides a robust RESTful API. This interface accepts JSON input and returns JSON output, making it compatible with any system capable of making HTTPS calls, including Python scripts, serverless functions, and complex enterprise workflows.

The API follows a standard authentication flow. Developers must first register their agent by providing a name and specifying how they wish to receive responses. The invocation parameter is critical here, as it determines whether the agent will use a persistent connection—where it actively fetches answers or maintains a WebSocket—or a webhook model, where Taskriz posts responses to a designated HTTPS endpoint.

{
  "name": "Invoice Approver",
  "invocation": {
    "mode": "persistent"
  }
}

Upon registration, the system issues a unique agentId, clientId, and clientSecret. The secret is displayed only once and must be stored securely. These credentials are then exchanged for a bearer token that lasts for one hour, ensuring secure, time-limited access to the API.

Pairing Agents with Humans

A defining feature of the Taskriz platform is its secure pairing mechanism, which ensures that AI requests reach the correct human operators. This is achieved through a temporary, one-time pairing code that is valid for 24 hours.

When an agent initiates a pair request, Taskriz generates a manual code, a pairing URL, and a QR code image. The human recipient can scan the QR code with their smartphone camera or click the link, which opens Taskriz with the pairing details pre-filled. Alternatively, they can manually enter the code in the Taskriz interface.

Only individuals who have explicitly paired with an agent can receive requests from it. This creates a secure, permission-based network where humans can also unpair at any time, maintaining control over their availability.

Once paired, the agent receives a paired event containing the human’s email and connection details. This establishes a trusted channel for sending high-stakes requests, such as financial approvals or legal reviews.

Designing Requests as Interfaces

Taskriz treats requests not just as text messages, but as interactive interfaces. The payload field in the API allows developers to define custom UI components using a syntax similar to Smalltalk’s Morphic framework. This enables the creation of buttons, labels, and forms that guide the human responder toward specific, structured outputs.

For example, an invoice approval agent might send a request with a summary of the vendor, amount, and due date, alongside "Approve" and "Query" buttons. When the human clicks a button, the agent receives a structured JSON response, such as { "decision": "approve" }, tied back to the original request via a correlationId.

This correlation ID is essential for maintaining context. It ensures that when the human response arrives—whether seconds or days later—the agent can correctly associate the feedback with the specific operation it was performing. The payload can range from simple text prompts to complex multi-step interfaces, allowing developers to tailor the human interaction to the complexity of the task.

By combining secure pairing, flexible invocation modes, and rich interface design, Taskriz’s Human-in-the-Loop API provides a comprehensive solution for adding human judgment to AI-driven processes.