Tencent Yuanbao Search Opens to External Users via Skill, MCP, and API

Tencent's internal "AI search recipe" has been made public. ZhiDongXi Download the app for exclusive early access. October 21, 2026, 20:30:15 From Beijing Updated October 21, 2026, 20:33:26

In Google's traditional search, users receive a list of web page links; but in Yuanbao Search, the same question presents results that are quite different.

And when Agents capable of calling tools and executing tasks become the focus, erroneous or outdated information in retrieval may be passed down along subsequent tool calls, affecting the judgment and execution of the entire task.

Search therefore not only needs to be comprehensive and timely in updates, but also as accurate as possible, and organize information into results that Agents can use directly.

For this fundamental problem, RAG (Retrieval-Augmented Generation) is one of the important technical paths to mitigate knowledge lag and improve answer reliability.

But developers and enterprises that want to build their own RAG need to establish and maintain long-term multiple links such as source collection, data cleaning, and index construction, which also means high costs, and the value of ready-made search APIs is precisely to lower the threshold for integration and operations.

Over the past year, Yuanbao Search has continued to optimize, integrating public resources across the web and Tencent's high-quality content, building an index scale of hundreds of billions, and has been tempered through Tencent internal applications such as WorkBuddy and ima.

The significance of this upgrade is not just the release of a search capability, but also that Tencent is opening up to the entire industry search infrastructure that has been refined through internal business and validated at scale, becoming a general service that external developers can call directly.

For developers, there are two options for adding search capabilities to an Agent: building the retrieval pipeline yourself, which allows more flexible control over search logic but also requires continuous investment in R&D, source maintenance, and operations; or integrating a ready-made search service, which can reduce this part of the engineering investment, while search quality and service stability depend more on the service provider.

After this upgrade, Yuanbao Search provides three integration methods: Skill, MCP, and API, and offers a limited-time free quota: 2,000 calls per month, and in October and November, up to 6,000 calls per month via invitation code. Developers can first integrate and test, then decide whether to use it for actual business.

ZhiDongXi conducted an actual test of Yuanbao Search via API. The integration process is not complicated. First, obtain an API Key on Tencent Cloud, then open the terminal and enter a curl command to initiate a search request:

-H "Content-Type: application/json; charset=utf-8" \

-H "Authorization: Bearer Secret Key" \

-d '{"Query":"Today's AI news"}'

Replace the "Secret Key" in the code with the actual API Key obtained, and after execution you can submit a query to Yuanbao Search and view the returned results in the terminal.

▲ Raw return results in the terminal

After completing the API call, we encapsulated the interface into a Skill and integrated it into DeepSeek Harness, allowing the Agent to autonomously call search according to task needs. In this way, search capability changed from a manual request to a link that can be called at any time in the Agent workflow.

After integration was completed, we tested basic news retrieval, for example entering a prompt to use Yuanbao Search to check today's social news, and we could see that Yuanbao returned results in only 17 seconds.

Ordinary web search results center on titles, summaries, and links. Next, the Agent can also combine web scraping tools to read relevant content and further complete information organization.

Yuanbao Search returns not only web results centered on titles, summaries, and links, but also a type of vertical source that provides structured data for scenarios such as weather, stock prices, and gold prices.

For example, querying "Beijing weather today" returns a structured VR card that reflects the information very clearly.

We also tried searching for Tencent's stock price, which likewise hit a vertical source. The results returned by Yuanbao Search included a stock price card, with data from Tencent Zixuan Gu.

This means that for tasks with a clear data structure, the Agent does not always need to first open a web page, scrape the main text, and then extract the required fields; it can directly process the data returned by the search interface.

For an Agent that needs to continuously call tools, organize information, and execute the next step, one less scraping and parsing step can improve efficiency considerably.

Subsequently, we selected tasks such as policy queries, academic research, and developer resource organization, and combined them with the scraping tools in DeepSeek Harness to further test Yuanbao Search's performance in actual scenarios.

We first tested policy information retrieval, entering "Use Yuanbao Search to query this year's AI-related policy information." After about 42 seconds, the Agent compiled a list of policy information, making a table by time, document name, issuing body, and key content, and further summarized several main policy threads.

From the results, the information mainly came from websites of authoritative institutions such as the Ministry of Industry and Information Technology and the Ministry of Commerce, which not only facilitates quick understanding of policy developments but also provides clear entry points for subsequently finding the original texts and verifying specific provisions.

Academic research is another scenario well suited to collaboration between search and Agents. Taking the recently discussed topic of recursive self-improvement (RSI) as our theme, we called Yuanbao Search to find related papers from the past six months and organize the research progress. After about two and a half minutes, the system returned the relevant materials.

On this basis, we further combined the arXiv API to obtain paper information and cross-check the literature, generating an RSI review. Working together, the two allowed the Agent to continue turning search results into more systematic research materials.

Developer resource compilation can also reflect Yuanbao Search's capabilities. We had DeepSeek Harness call Yuanbao Search to find Python beginner learning resources, and it ultimately compiled 67 items, progressing step by step from environment setup and basic syntax to exercises and project practice, and distinguishing official documentation from community tutorials, forming a clearly structured learning list.

The value of this type of task lies not only in finding web pages. For developers, the search service provides the information entry point, while the Agent is responsible for organizing the information into results that better fit actual needs.

Finally, we further increased the task complexity, having DeepSeek Harness call Yuanbao Search to sort out Tencent's recent product developments, personnel changes, official technical information, and key business data, and generate a resource list with source links.

The final output was close to 300 lines, divided into four parts, including dates, event summaries, and source links, with some financial information further detailed down to metrics such as quarterly revenue and capital expenditure. The system also graded some sources and provided warnings for information lacking official confirmation.

From a single news search, to policy information aggregation, academic material organization, and then to enterprise-focused research, these test groups present a relatively clear path:

Yuanbao Search is responsible for helping the Agent discover and obtain external information, and the Agent then organizes this information into tables, learning lists, or topic-specific materials according to task requirements.

It should be noted that the task completion times mentioned above are the overall time taken by the Agent to execute the task, not the response time of the search interface itself, and include processing steps such as the model organizing results and subsequent web page crawling.

During our testing through VS Code, with five consecutive calls, the API took 345–485 milliseconds per call, averaging only about 0.42 seconds. The speed is quite fast, and in practice it feels almost like an instant reply after the instruction is issued.

WorkBuddy needs to search external information to support office tasks such as research and analysis; ima needs to combine internet information with the user's existing knowledge; Tencent Video and Tencent Financial Technology involve content retrieval and information queries in professional fields, respectively.

In WorkBuddy, Yuanbao Search can help the agent understand search intent more accurately, find the needed materials faster in scenarios such as technical documentation and official APIs, and reduce unnecessary tool calls and model rounds.

The WorkBuddy team reported that in the past, doing research required first collecting a large number of web pages and then judging one by one whether the information was credible. After integrating Yuanbao Search, authoritative information can directly enter the workflow, and conclusions can also be traced back to their sources.

The team therefore no longer needs to maintain its own information sources and crawling pipelines, and can devote more energy to task planning and delivery for the agent.

According to data disclosed by Tencent, Yuanbao Search has reached hundreds of millions of daily calls, supporting more than 50 core businesses including WorkBuddy, ima, Tencent Financial Technology, and Tencent Video.

Ma Jin, Vice President of Tencent Cloud and head of R&D for Tencent Yuanbao Search, said in an interview with media including Zhidongxi that in WorkBuddy, after integrating Yuanbao Search, the average number of search calls per task dropped from about 8 to about 5, and Yuanbao Search's internally evaluated search satisfaction also improved by 10%–20% compared with before the upgrade. The response time of the fastest version of the interface was optimized from the original 400–500 ms to within 300 ms.

These internal applications have provided experience using the search capability in different business scenarios. However, validating internal products and providing standardized services externally are not the same thing.

Internally, search can be adapted around established models, data, and business processes; for external developers, compatibility with different models, Agent frameworks, and application scenarios must be considered, and developers must also be able to call it independently, combine it flexibly, and rely on stable service quality.

This is also one of the focuses of this Yuanbao Search upgrade. Through Skill, MCP, and API, Tencent packages the search capability into different access methods for developers to use as needed.

As large models begin to access information online, search is becoming an important component of models and Agents. However, vendors have not all adopted exactly the same product approach.

Take OpenAI and Google as examples: both integrate search into their model APIs. OpenAI retrieves information online through the Web Search tool, while Google calls search through Google Search grounding and uses the retrieval results and source information to generate answers.

It should be noted that this type of "model built-in search" does not mean that the model weights themselves possess search capability; rather, the platform connects retrieval, result processing, and answer generation.

Another path is to open search as an independent service, so developers can call the search capability separately, then choose a model according to business needs, and connect the retrieval results to the Agent workflow.

The difference between the two paths is that the former integrates search with model invocation, reducing the work developers must orchestrate themselves; the latter makes search capability independent, giving developers more room for model choice and process control.

Therefore, for an independent search API like Yuanbao Search, the key to competition is no longer just whether information can be found, but also source quality, response speed, invocation cost, and whether it can stably support an Agent's multi-step retrieval and task execution.

From this perspective, this upgrade of Yuanbao Search should not be judged only by how many access methods it provides, but also by whether its search capability can continuously provide useful information in different business scenarios.

According to information disclosed by Tencent, Yuanbao Search has built a hundred-billion-level index, covering 13 vertical industries such as finance, government affairs, and healthcare, and has connected 157 authoritative sources.

At the retrieval technology level, Tencent said it uses a self-developed disk vector index integrated with an inverted index; the related optimization raised recall from below 80% to over 93%, while reducing the number of machines from 3,200 to 400.

After finding information, the system also needs to determine which content is more worth presenting. Yuanbao Search reconstructs the ranking pipeline through large models and uses Gist Token compression technology to optimize inference latency and cost, seeking to balance result quality and service efficiency.

In terms of source governance, Ma Jin mentioned in an interview that Yuanbao Search evaluates content quality at three levels—site, account, and page—and judges source value by combining factors such as authenticity, content richness, and professionalism. For different questions, it also adjusts the weights of authority and timeliness.

These technical indicators show Tencent's investment in retrieval, ranking, and content governance, but they cannot be directly equated with the accuracy of final answers. For developers, what matters more is whether these capabilities can continuously translate into reliable search results in actual tasks, as well as a stable and controllable invocation experience.

Behind this is Tencent's judgment on the Agent infrastructure market. As AI applications move from answering questions to executing tasks, developers' demand for real-time information, reliable sources, and stable retrieval capabilities is also increasing.

Tencent hopes to package its internally accumulated search capabilities into out-of-the-box cloud services through an independent AI search service, so that developers and enterprises can add search capabilities to their AI applications faster and at lower integration cost without building retrieval pipelines from scratch.

Conclusion: Tencent places another piece in the B-end Agent market

As Agents move from answering questions to executing complex tasks, competition in the AI industry is also extending from model capabilities to more complete infrastructure and developer ecosystems.

For cloud vendors, whoever can provide developers with reliable information sources, convenient tool integration, and stable service support will be more likely to become an important foundation for the deployment of Agent applications.

"Special statement: The above content (including videos, pictures, or audio if any) is uploaded and published by users of "Dafeng Hao," a self-media platform under Phoenix News. This platform only provides information storage space services.

Notice: The content above (including the videos, pictures and audios if any) is uploaded and posted by the user of Dafeng Hao, which is a social media platform and merely provides information storage space services."