The Search Layer for all your context needs
One search layer for what your agent remembers, what's public, and what's yours. Memory, Public Search, and Proprietary Search
Use cases
One search layer.
Three kinds of context.
One layer for what your agent remembers, what's public, and what's yours — each returned as a cited, context-ready packet.
What your agent knows
·
What the world knows
·
What you know
Memory · what your agent knows
Your agent's memory, queryable in milliseconds.
Fast
Scoped
Cited
Small, self-generated facts your agents write as they work — scoped per project and per agent, returned with the source they came from.
Self-written facts
Written as the agent works, recalled in milliseconds.
An agent writes a fact once; any scoped agent reads it back instantly, with the source attached.
Write
"Retry failed requests up to 3× with exponential backoff before surfacing an error."
→
FACTconf 0.97
8 ms
"Retry failed requests up to 3× with exponential backoff…"
Scoping
Isolated per project & agent.
Facts never leak across projects — each is scoped to who's allowed to read it.
checkout-service
3 can read
billing-api
2 can read
data-export
1 can read
Provenance
Every fact carries its source.
No unattributed memory — trace any fact back to where and when it was written.
"Retry failed requests up to 3× with exponential backoff…"
↳ cited 4× across sessions · 2h ago
Public Search · what the world knows
The open web, indexed for agents.
Cheap
Authority-ranked
Fresh
Authority-ranked, fresh results across multiple indexes — built for context windows, not browsers.
Authority-ranked
Ranked results, extracted to clean markdown.
Ninelayer ranks the open web by authority, then hands your agent the extracted content — not a list of links to crawl.
Extract · markdown247 ms
# Caching in Next.js
Dynamic APIs opt routes out of the
full route cache by default.
- Router Cache · Data Cache
Dynamic APIs opt routes out of the
full route cache by default.
- Router Cache · Data Cache
1,240 tokens returned5× cheaper than raw search
Token cost
5× cheaper per query.
Raw search API6,400 tok
Ninelayer1,240 tok
Freshness
Fresh, never frozen.
Indexes refresh continuously, so agents reason over the current web.
Proprietary Search · what you know
Search your corpus, without losing control.
Secure
Isolated
Predictable
Customer-uploaded data, indexed at enterprise scale — isolated, encrypted, and never used for training.
Your corpus
Your uploaded data, indexed and isolated.
Point Ninelayer at your docs, wikis, and channels; every answer stays inside your tenant.
Indexed sources18,400 chunks
INTERNALconf 0.99
"Rotate service tokens every 90 days via the internal vault CLI."
source: docs.acme.internal/security/token-rotation
Security
Isolated, encrypted, never trained on.
✓Full tenant isolation, always
✓Encrypted end-to-end (TLS + at rest)
✓Never used to train a model
✓Instant delete, no soft-deletes
Pricing
Predictable at scale.
Flat rate
$0.40/ 1k chunks indexed
18,400 chunks≈ $7.36 / mo
✓ no per-query fees · no surprises
Technical
Tavily and Exa return a URL, a relevance score, and a content blob — the same shape as a raw search result. Ninelayer returns typed evidence packets: every result is classified by source type (official_docs, github_repo, community_evidence), attributed to a framework, assigned an authority tier (primary vs supporting), and scored with a confidence value. Your agent knows what kind of source it is reading before it reads a word. No custom parsing, no post-processing rerankers, no heuristics to write.
Standard search APIs return links, messy HTML, and SEO-ranked pages. Ninelayer returns authority-aware evidence packets: primary sources, supporting context, and community evidence in compact Markdown built for a context window.
An evidence packet is a compact search response designed for agents. It separates primary authority sources from supporting evidence and community context so the agent knows what to trust first.
No. Ninelayer is the search layer for AI agents. We are starting with coding agents because their failures are easy to measure: wrong docs lead to wrong code. The same model applies to research, support, sales, ops, and browser agents.
Sign in, create an auth token from the dashboard, and connect Ninelayer to your MCP-compatible agent or custom workflow.
A token is required for agent workflows. You can explore recipes and examples without one, then create a token from the dashboard when you are ready to connect your agent.
Anything that speaks the Model Context Protocol (MCP). This natively includes Claude Code, Cursor, Windsurf, and any custom agent built on frameworks that support MCP tools.
No. Ninelayer handles retrieval, extraction, source grouping, and formatting. The data is fed directly back to your agent as a clean string of facts, ready for immediate execution.
We index the live web with a heavy bias toward current developer documentation, API references, migration guides, release notes, GitHub issues, and technical forums to reduce stale-code hallucinations.
Yes. Ninelayer is a drop-in alternative to Tavily for AI agent workflows. Where Tavily returns raw search results, Ninelayer returns authority-ranked evidence packets pre-formatted for agent context windows — at roughly one-fifth the cost per query on the Pro plan.
Yes. Exa is optimized for semantic similarity search and returns raw content. Ninelayer is purpose-built for agentic workflows: it returns structured evidence packets with source-type classification, authority scoring, and compact Markdown output — consuming fewer tokens while delivering sufficient grounding for coding, research, and support agents.
Serper and Brave Search return raw SERP data — titles, URLs, and snippets intended for display. Ninelayer post-processes and ranks results specifically for LLM consumption: content is extracted, deduplicated, typed by source category, and returned as compact Markdown. Agents receive grounding data, not a search results page.
Ninelayer provides three core tools: ninelayer_deep_search for live web research across docs, GitHub, packages, and the web; ninelayer_get_url for extracting clean Markdown from a specific URL; and academic_search for paper-level search over the arXiv index. All tools are available over MCP (mcp.ninelayer.in) or a plain stateless HTTP API at /mcp/tools-call.
Run: claude mcp add ninelayer --transport http --url https://mcp.ninelayer.in/mcp/ and set your NINELAYER_AUTH_TOKEN environment variable. Your Claude Code sessions will then have access to ninelayer_deep_search and ninelayer_get_url automatically.
In Cursor settings, add a new MCP server with transport HTTP and URL https://mcp.ninelayer.in/mcp/. Pass your auth token as the NINELAYER_AUTH_TOKEN header or environment variable. Cursor will surface Ninelayer tools to any agent session automatically.