$26M Seed Funds Keenable's 100B-Document AI Index

Keenable has exited stealth with a $26 million seed round and an independent web index built for AI models and autonomous agents, as detailed in its August 25 announcement. The San Francisco startup says its index contains more than 100 billion documents and returns results in under 250 ms at p95 in US East. For teams building production agents, the pitch is direct access to current web information without depending entirely on search infrastructure controlled by large consumer platforms.

Funding and Founding Team

Accel led the financing, with participation from Conviction Partners and angel investors associated with Amazon, ClickHouse, Databricks, Google, Hugging Face, NVIDIA, Snowflake, and SpaceX. Zhenya Loginov, a partner at Accel, led the investment for the firm.

Keenable was founded in 2025 by Andrey Styskin and Matthias Petri. Styskin previously served as CEO of Search and Ads at Yandex, where he built a 200-billion-document web index, and as Director of Web Infrastructure at Amazon AGI. Petri was a Principal Applied Scientist at Amazon AGI and built the web grounding service behind Alexa.

The founders are targeting a retrieval problem created by agent workloads. Human search systems can return ranked pages and summaries for people to inspect. Agents need structured access to source documents, low response times across many tool calls, and pricing that remains viable when retrieval happens repeatedly during a multi-step task.

Search Infrastructure for Agent Runtimes

Keenable provides access through standard APIs, a command-line interface, and a Model Context Protocol (MCP) server at https://api.keenable.ai/mcp. MCP support makes the service accessible to agent runtimes that already use the protocol for tool discovery and execution. Developers working through MCP fundamentals can treat the search layer as an external tool rather than embedding a provider-specific integration into every agent.

The company is also developing a Time Machine API. Agents can specify a query_time parameter to search earlier snapshots of web pages, with both the document corpus and historical rankings rewound to the selected point in time. This supports tasks such as reconstructing prior product documentation, validating what information was available during an event, or training systems against historical web states.

A Web Query Language (WQL) is in development for querying, retrieving, and cross-referencing thousands of live web sources simultaneously. The design points toward retrieval workloads that exceed a conventional search box, particularly research agents that need to gather evidence across many domains before producing an answer.

Pricing and Availability

Keenable offers a 100,000-request trial grant and two commercial tiers.

Tier	Intended users	Deployment	Price
Agent Builder Tier	Agent builders and enthusiasts	Cloud-only	$4 per 1,000 requests
Frontier Tier	AI labs and inference platforms	Cloud and on-premises	$1 per 1,000 requests at 100+ RPS

The pricing distinction reflects two different workloads. A developer testing an agent may prioritize simple cloud access, while an inference provider needs sustained throughput and the option to run on premises. If your system performs several searches per task, request volume should be modeled alongside model inference and token costs, using the same discipline applied to production RAG systems.

Voice Agents and Production Retrieval

Keenable says its API is already used in production by several undisclosed AI labs and inference providers for offline training and live retrieval. The company has also partnered with Gradium to add real-time web search to voice agents built with the open-source Gradbot framework.

For spoken interfaces, the under-250-ms p95 target matters because search latency joins speech recognition, reasoning, and text-to-speech in one conversational loop. The Gradium integration connects this infrastructure to real-time voice agents, where a slow lookup can interrupt turn-taking even when the model responds quickly.

Keenable enters a retrieval market that includes Tavily and Exa as developers seek alternatives to restricted or increasingly expensive search APIs. Teams evaluating the service should measure answer quality, freshness, source coverage, historical reproducibility, and end-to-end latency in their own agent traces. The practical decision is whether an independent index can reduce retrieval cost and dependency while meeting the evidence requirements of your workload.