A 100-billion-document index built for AI agent searches

Keenable raised $26 million to give AI agents fast web search, MCP access, and results from the web as it existed at a specific time.

Agents that perform tasks regularly need to search the web to supplement a model’s knowledge, verify recent information, or locate specialized sources. Keenable wants to provide that search layer without relying on results pages designed for humans to browse.

The startup has emerged from stealth with a $26 million seed round. Accel led the investment, with participation from Conviction Partners and individual investors from companies including Google, xAI, Nvidia, Amazon, and Hugging Face. Keenable did not disclose its valuation, each participant’s contribution, or the financial terms of the deal.

Keenable was founded by Andrey Styskin and Matthias Petri. Styskin previously led Yandex’s search and advertising businesses before working on web infrastructure at Amazon AGI. Petri was a principal applied scientist at Amazon, where he helped build the web search service used to ground some Alexa responses in online information.

The funding will support a team that currently includes around 15 engineers across the United States and Europe. The company plans to double its headcount by the end of the year and develop its go-to-market operations while continuing to fund the crawling, storage, and updating of its index.

Keenable says it has already indexed more than 100 billion documents. That would give it a substantial corpus among search providers serving AI agents, but the figure comes from the company and has not been independently audited. Size alone does not determine search quality. Page freshness, geographic coverage, duplicate-content handling, resistance to spam, and ranking relevance are equally important.

Its first product is a search API. Queries can be restricted to a specific website, publication period, or acquisition period within Keenable’s index. Responses include the title, URL, description, excerpt, publication date, and indexing date. A second operation retrieves the contents of a selected page so an agent can read, summarize, or use it as a source.

Separating search from content retrieval avoids transferring entire documents before determining which results are relevant. This structure better reflects how an agent works: it can run several searches, review excerpts, select a few results, and then request the text of the chosen pages.

Access is available through a REST API, command-line interface, and MCP, the protocol that allows an assistant to discover and call external tools. Keenable’s MCP server exposes its search and content-retrieval functions to environments such as Claude, ChatGPT, Codex, Cursor, and Windsurf. The company also documents integrations with LangChain, LlamaIndex, Haystack, Mastra, and the Vercel AI SDK.

This compatibility does not mean Keenable runs the models themselves. The service supplies results and documents to the agent making the request. Reasoning, synthesis, and response generation remain the responsibility of the model and development environment selected by the user.

The second product, Time Machine, adds a historical dimension. By including a `querytime` parameter, an agent can search the index as it existed at a specified date. Pages acquired after that point are excluded, and results are ranked using the corpus that was available at the time.

This can help reconstruct what information was available before an event, compare changes in a company’s messaging, or prevent historical research from being influenced by articles published after the fact. An agent investigating an acquisition, for example, could request results available before the announcement and compare them with those that appeared afterward.

Time Machine should not be confused with a complete copy of the web. It can only retrieve pages Keenable crawled, recorded, and retained. A page missing from its index at the requested date will remain unavailable even if it existed at the time. The feature also does not promise to reproduce a website’s full layout, media, or interactive elements exactly as they appeared in a browser.

Keenable reports p95 response times below 250 milliseconds from the AWS US East region. According to its own measurements, this means 95% of the relevant requests are completed within that threshold under the tested configuration. The figure is not a universal guarantee: distance from the server, query type, system load, result volume, and the subsequent retrieval of individual pages can all increase the total delay experienced by an agent.

This speed is particularly relevant to voice applications, where even a few seconds of silence can disrupt a conversation. Keenable has demonstrated an integration with Gradium’s voice-agent framework. The partnership provides one public example of the technology in use, but the company has not identified the other AI labs and inference providers it says are already using its infrastructure in production, during training, or at runtime.

The service currently includes 100,000 free requests per month. Beyond that allowance, the agent-builder plan costs $4 per 1,000 requests and runs in Keenable’s cloud. The advertised price of $1 per 1,000 requests applies to a dedicated-capacity plan starting at more than 100 requests per second, intended for AI labs and inference platforms. That version can run either in the cloud or on the customer’s own infrastructure.

The $1 figure therefore does not represent the standard cost for a small project. It applies to a level of volume and capacity that requires a more structured commercial agreement. An agent’s actual operating cost will also depend on the number of searches initiated for each task, the pages retrieved, and the model calls required to process the results.

To compare its service with Brave, Exa, Parallel, Perplexity, Serper, Tavily, and You.com, Keenable publishes NEEDLE, an open benchmark focused on search for AI agents. Its tests cover five categories: specialized literature, breaking news, company information, rare entities, and legal documents.

Each provider receives the same queries.