Open Agentic Kit overview (closed beta)
Open Agentic Kit (OAK) is The Trade Desk's MCP-based access to Ask Koa, which brings platform data and actions directly into the AI systems your traders and operations teams use.
NOTE: OAK access is currently available to select partners in closed beta. To register your interest, contact your account representative at The Trade Desk.
Benefits
Whether your traders review campaign pacing, investigate underperformance issues, or have an autonomous agent run optimizations around the clock, OAK transforms complex operations into simple conversational queries and removes the friction:
- No platform UI to navigate.
- No broken workflows from switching interfaces and copying data.
- No custom API integration to build or maintain.
- No vendor lock-in. OAK works with AI assistants your teams already use.
How it works
Rather than exposing a list of API endpoints, OAK allows your client assistant access to the Ask Koa tool, and send prompts about how the platform works: the workflows, the entity relationships, the context behind decisions, and the best practices that drive results. The Koa Agent system reasons about your goals, handles multi-step and cross-entity tasks, and adapts as your use cases evolve—and because that intelligence lives in the workflow layer, new platform capabilities become available automatically as they ship, with nothing to redeploy on your end. It relays all relevant information back to your client assistant and offers to make changes to your account.
OAK is built on the Model Context Protocol (MCP), the open standard supported in Claude, ChatGPT, Gemini, Cursor, and most AI-powered applications. If your organization has already invested in AI infrastructure, The Trade Desk connects to it today with a single configuration entry.
Capabilities
The following table outlines what you can do with OAK today, and what's coming later this year.
| Capability | Release status | What you get |
|---|---|---|
| Knowledge Search | Available | Search The Trade Desk's documentation and knowledge base in real time. Your AI assistant gets accurate guidance on platform setup, campaign structure, targeting options, optimization best practices, and troubleshooting—grounded in live platform knowledge, not stale training data. |
| Data Retrieval | Available | Pull live and historical data across campaigns, ad groups, creatives, audiences, seeds, and deals—pacing metrics, performance data, and audience insights. Refine queries conversationally without needing to know the exact API call. Your AI assistant maintains context across the conversation, so you can drill down, pivot, and act without starting over. |
| Configuration & Management | H2 2026 | Create and modify core entities in your account—campaigns, ad groups, creatives, audiences, and deals—and optimize targeting configurations for improved performance. All changes follow a multi-step flow: your AI assistant proposes the modification, you review and confirm, then it executes. The confirmation is enforced at the platform level—no surprise modifications, whether operated by a human trader or an autonomous agent. |
| Reporting | Available | Retrieve large volumes of performance data in seconds and compare trends across time periods—for example, an advertiser's performance over the last 7 days versus the prior 7 days. Reports come back structured for immediate analysis and action. |
| Insights | H2 2026 | AI-driven analysis of campaign performance powered by Koa, The Trade Desk platform intelligence: where your budget is working, where it isn't, and what to do about it. |
| Troubleshooting | H2 2026 | Identifies campaign issues and surfaces actionable fixes powered by Koa. Your AI assistant can resolve them immediately using Configuration & Management—closing the full detect-to-fix loop without any manual handoff. |
NOTE: Knowledge Search, Data Retrieval, and Reporting are read-only and safe to enable broadly. Configuration & Management requires a write-scoped token. For details, see Authentication methods.
Example prompts
Here are a few examples of what you can ask OAK to do.
| What you ask | What happens |
|---|---|
| "What are the best practices for setting bid factors?" | Searches The Trade Desk knowledge base and returns platform-specific guidance in context. |
| "Get me all campaigns underpacing their daily budget." | Returns a list with current pacing metrics so you can act immediately. |
| "Show me spend and pacing across all campaigns for advertiser X." | Returns a structured summary with options to drill down by ad group, filter by date range, or take action directly. |
| "Show me my advertiser's performance over the last 7 days compared to the prior 7 days." | Retrieves both periods in seconds and returns a side-by-side trend comparison—surfacing what moved, by how much, and where to look next. |
| "Set the total budget to $5,000 with a $250 daily cap for flight X on campaign Y." | Proposes the updated budget and daily cap for your confirmation, then applies them to the flight. No manual follow-up needed. |
| "Pause all display ad groups for advertiser X." | Identifies the matching ad groups, then upon your confirmation pauses all of them. No manual follow-up needed. |
Ready to connect? See Get started with Open Agentic Kit.
AI terminology
These sections define terms used throughout OAK documentation, including general AI terminology and The Trade Desk terminology.
General terms
The following table lists introductory terms for a general understanding of AI.
| Term | Definition |
|---|---|
| Agent | An AI system that perceives, reasons, and acts to achieve goals in an environment. |
| Artificial Intelligence (AI) | An umbrella term for fields that include: machine learning, natural language, computer vision, multi-agent systems, and more. |
| Code LLM | An LLM specialized in generating code using generative AI. |
| Generative AI | An application of machine learning that specializes in synthesizing text, images, audio, or video. |
| Large Language Model (LLM) | A generative AI model trained on data to understand prompts and generates an output. |
| MCP | The Model Context Protocol protocol that allows client assistants to interact with tools. |
| Multi-agent | A network of multiple agents interacting with each other to achieve individual or shared goals. |
| Multi-modal LLM | An LLM that accepts other forms of inputs beyond text, such as image, audio, or video files. |
| Skill | A set of instructions that are designed to accomplish a goal. Skills can range from running a script to planning and executing modifications using an LLM. |
| Tool | An endpoint intended for use by client assistants to interact with other AI assistants. |
The Trade Desk terms
The following table lists terms specific to OAK and Koa.
| Term | Defintition |
|---|---|
| Ask Koa | A tool in OAK exposed as an endpoint that enables communication between client assistants and Koa, an AI created by The Trade Desk. |
| Capabilities | A label for a set of tasks performed by Koa. Tasks are completed by Koa Agent using internal skills made to achieve platform-specific goals. |
| Client assistant | Your MCP-compatible AI client that communicates with OAK. |
| Koa | The AI that powers The Trade Desk platform. |
| Koa Agent | The Trade Desk AI system that handles answering prompts from other agents. |
| Koa assistant | The Trade Desk MCP-compatible AI client available in the Kokai UI. |
| Open Agentic Kit (OAK) | OAK is The Trade Desk's MCP-based access to Ask Koa. |