Keenable
by Keenable
Independent web search API for AI agents — 100B+ document index, sub-250ms p95
Keenable is an independent web search API for AI agents, built on its own crawler, index and ranking stack rather than a reseller feed from Google or Bing. It targets AI labs, inference providers and agent developers who need grounded, low-latency retrieval, and offers a 100-billion-document index, sub-250ms p95 latency and published per-request pricing.
Keenable is an independent web search API built specifically for AI agents and language models rather than for human browsers. The company exited stealth on 25 August 2026 with a $26 million seed round led by Accel, with participation from Conviction and angel investors who work at Google, xAI, NVIDIA, Amazon and Hugging Face. It was founded by Andrey Styskin, who previously led search, AI and cloud at Yandex, and by AI scientist Matthias Petri, and it runs on a team of roughly 15 engineers that the founders plan to double by the end of 2026. Rather than reselling Google or Bing results, Keenable operates its own crawler, index and ranking stack, which it says spans more than 100 billion documents and returns results in under 250 milliseconds at p95 from US East. The product offers three integration paths: a REST endpoint at api.keenable.ai/v1/search authenticated with an X-API-Key header; an MCP server that Claude Code, Claude Desktop, Cursor and Windsurf can call directly, including a keyless public tier with an hourly request cap; and a CLI that auto-configures MCP for whichever clients it detects. Alongside search it exposes a fetch_page_content tool that returns pages as clean markdown, an early-access Time Machine product, and a planned Web Query Language for composing answers across multiple web sources. Pricing is published and usage-based: $4 per 1,000 requests on the Agent Builder tier, and $1 per 1,000 requests on a Frontier tier available cloud or on-premises for AI labs sustaining 100+ requests per second.
The platform or applied-AI team that already pays a search API bill for agent grounding and is being squeezed on latency or per-query cost — Keenable's published $4 per 1,000 requests undercuts the $5-7 range competitors advertise, and the Frontier tier drops to $1 per 1,000 at sustained volume.
Grounded web retrieval for agents at roughly a quarter of typical list price, from an index the vendor owns end to end rather than resells, with an on-premises option for labs that cannot send queries to a third party.
At a Glance
- Category
- Agent Development
- Pricing
- Usage-based, Freemium
- Target Market
- CTOs, Enterprise Developers, Data Scientists, AI Engineers
- Deployment
- API-based, Cloud-first, Self-hosted
- Team Size
- 11-50
Key Features
- ✓Owned crawler, index and ranking stack
Keenable crawls and ranks its own corpus of 100B+ documents rather than reselling Google or Bing results, which removes a supply-chain dependency that has repeatedly forced repricing on SERP-scraping competitors.
- ✓Sub-250ms p95 search latency
Published p95 under 250 milliseconds from US East, with a commenter on the Show HN thread reporting p50 under 200ms — material when an agent loop issues several searches per user turn.
- ✓MCP server with a keyless public tier
Agents can call Keenable through Model Context Protocol with no account at all, subject to an hourly cap; adding an API key lifts the limit, so evaluation costs nothing but a config line.
- ✓fetch_page_content tool
Retrieves an arbitrary URL and returns it as clean markdown, so an agent can go from a search hit to full readable page text without a separate scraping vendor.
- ✓CLI that auto-configures agent clients
One command detects and wires up MCP for Claude Code, Claude Desktop, Cursor and Windsurf, which removes the hand-edited JSON config step that breaks most MCP onboarding.
- ✓Frontier on-premises tier
AI labs sustaining 100+ requests per second can run Keenable in their own environment at $1 per 1,000 requests, so training-scale retrieval does not have to leave their infrastructure.
- ✓Time Machine and Web Query Language
Time Machine is in early access for historical web state, and a SQL-like Web Query Language is planned to let agents compose answers across several sources in one call.
Capabilities
Use Cases
- •Grounding a customer-facing agent
An enterprise assistant issues a Keenable search per user question and cites the returned URLs, so answers about current pricing, outages or regulation are sourced rather than hallucinated.
- •Coding agent documentation lookup
Claude Code or Cursor calls the MCP server to pull current library docs and changelogs before generating code, which cuts errors caused by a stale model knowledge cutoff.
- •Training-time data collection at an AI lab
A lab runs the Frontier on-premises tier at 100+ requests per second to gather and refresh web corpora without routing every query through an external vendor's cloud.
- •Cutting an existing search API bill
A team already spending on Exa, Tavily or Serper reroutes high-volume agent traffic to Keenable at $4 per 1,000 requests and measures quality against the public NEEDLE benchmark before committing.
- •Competitive and market monitoring agents
A scheduled agent searches for named competitors daily and uses fetch_page_content to pull each hit into markdown for summarisation into an internal digest.
Ideal For
Best For
- ✓Grounding RAG and agent answers in live web content with source URLs attached
- ✓AI labs and inference providers that need web retrieval during both training data collection and runtime serving
- ✓Coding agents in Claude Code, Cursor or Windsurf that need current documentation via a drop-in MCP server
- ✓Cost-sensitive high-volume search workloads where per-1,000-request price dominates the inference bill
- ✓Teams that need an index not derived from Google or Bing SERP scraping for licensing or resilience reasons
Not Ideal For
- ✗Enterprises searching internal documents, wikis or SharePoint — Keenable indexes the public web only and does no permission-aware retrieval over corporate content
- ✗Teams that need a proven, referenceable vendor today: Keenable exited stealth in August 2026, discloses no named customers, and its Show HN drew only 10 points and 5 comments
- ✗Buyers who rank raw result quality above price and latency — the independent Artificial Analysis Search Index scores Keenable at 67 against 68-74 for Exa variants and 75 for Parallel Search (advanced)
- ✗Regulated buyers who need SOC 2, ISO 27001 or a data-processing agreement in hand before signing, none of which the company publishes yet
Deployment
Market Analysis
Pros
- ✓Published, aggressively low per-request pricing in a category where most vendors quote on request
- ✓Independent index rather than a Google/Bing SERP reseller, which removes a common single point of failure and repricing risk
- ✓Keyless MCP tier means an agent team can evaluate it in minutes without procurement or a credit card
- ✓Founding team has genuinely relevant pedigree — Andrey Styskin previously ran search, AI and cloud at Yandex
- ✓Runs NEEDLE, a public benchmark that scores it against named rivals daily, which is a rare transparency posture for a vendor
Cons
- ✗Independent scoring puts result quality behind the leaders: Artificial Analysis rates Keenable at 67 against 68-74 for Exa variants and 75 for Parallel Search (advanced)
- ✗Zero named customers — the company will only say 'several AI labs and inference providers', plus a partnership with voice AI firm Gradium
- ✗Fifteen-person team with a $26M seed against a cost base TechCrunch describes as 'painfully expensive', in a market where Google and Microsoft already own indexes
- ✗No published SOC 2, ISO 27001, GDPR/DPA or data-residency documentation, which will stall enterprise procurement
- ✗Negligible practitioner track record so far — the Show HN drew 10 points and 5 comments, and there is no Reddit, G2 or Capterra presence to check
Pricing
Public / keyless MCP
$0
- ✓MCP server access with no account
- ✓Hourly request limit
- ✓Search and fetch_page_content tools
Agent Builder
From $4 per 1,000 requests
- ✓Cloud, pay-as-you-go
- ✓Search API and fetch_page_content
- ✓MCP server and CLI
- ✓API key removes hourly cap
Frontier
From $1 per 1,000 requests
- ✓For AI labs at 100+ RPS
- ✓Cloud or on-premises deployment
- ✓Volume rate limits
Metered per request, not per seat, and unusually for this category the list price is published rather than gated behind sales: $4 per 1,000 requests on Agent Builder and $1 per 1,000 on the Frontier tier for labs sustaining 100+ RPS, where on-premises deployment is also available. A Show HN commenter framed the $4 figure against the $5-7 per 1,000 competitors advertise. Note that the independent Artificial Analysis benchmark reports far higher effective cost per completed agent task (roughly $63-73 per 1,000 tasks) because a single task issues many searches — budget against your own query-per-task ratio, not the headline rate.
Security & Compliance
Connect
Sources
This page was written from 7 sources, 6 on domains other than keenable.ai.
Stay Ahead of the Curve
Weekly enterprise AI insights for technology leaders. No spam, no vendor pitches—unsubscribe anytime.
SubscribeRelated Products
LangWatch
Open-source agent testing platform pairing simulation-based QA with OpenTelemetry tracing and evals
LangSmith
Framework-agnostic platform to observe, evaluate, deploy and continuously improve production AI agents
Composio
Managed tool-calling and authentication layer connecting AI agents to 1,000+ enterprise applications
Kitesurf
An agent-first browser that runs in V8 isolates on Cloudflare Workers