For everyone building on agents

Stop burning tokens
on WebFetch.

A 40-page site is ~80k raw-HTML tokens and six roundtrips by the time your agent parses the DOM. Lattis pre-indexes the web — chunked, embedded, cited, kept fresh. One endpoint, every site your agent touches.

Hosted MCP · REST API · npx @lattis-dev/cli

The actual cost

Two ways to read a 40-page docs site

Without Lattis

~80k

tokens of raw HTML per agent task

  • 4–8 round trips of WebFetch + DOM parse
  • Truncated context, stale answers, brittle scrapers
  • Token bills you can't budget

With Lattis

~2k

tokens for the same answer

  • One MCP call returns ranked chunks with citations
  • Pre-chunked, pre-embedded, deduplicated
  • Every result carries a freshness stamp

Numbers are characteristic, not measured against a formal benchmark — the qualitative shape is what matters at the install decision.

Try it without installing anything

Live query against the production index

This box hits api.lattis.dev/search from your browser. Same endpoint an agent would call. No signup, no demo data, no caching tricks.

For Agent Builders

One endpoint for every site, instead of one integration each

Pre-indexed, kept fresh, served through MCP, REST, or a CLI. Stop wiring up MCPs by hand. Stop falling back to WebFetch and parsing HTML on the agent's clock.

MCPClaude · Cursor · any client

https://mcp.lattis.dev/mcp

Cross-site server. Every indexed site is a tool. Each result carries a last_crawled stamp. Per-site endpoints at /s/<domain>/mcp.

API<200ms at the edge

POST api.lattis.dev/search

REST endpoints that return structured chunks, citations, and freshness metadata.

CLI--agent for JSON

npx @lattis-dev/cli

Terminal-native. Drop into any agent loop. Works offline of the MCP stack.

example cross-site query
$ npx @lattis-dev/cli search \
   "crm with free tier and webhooks" --agent
Returns ranked matches across every indexed CRM, each with a link to the live source page and a per-result freshness stamp.

Read the docs →

Wire it into your agent

Paste one block into your client

Remote where your client speaks streamable-http, npx where it doesn't. Same endpoint underneath. No credentials to manage.

Add to ~/Library/Application Support/Claude/claude_desktop_config.json

config
{
  "mcpServers": {
    "lattis": {
      "command": "npx",
      "args": ["-y", "-p", "@lattis-dev/cli", "lattis-mcp"]
    }
  }
}

Restart Claude Desktop after saving. On Windows the path is %APPDATA%\Claude\claude_desktop_config.json.

Something else? See the full docs →

How we compare

Pick the right tool for the job

Lattis isn't a web crawler. It's a pre-indexed agent-ready interface to the parts of the web your users care about.

LattisFirecrawlBrave Search API
Primary shapePre-indexed agent interfaceOn-demand crawler APIWeb search API
Hosted MCP serverYesYesNo
Per-site MCP endpointEvery indexed siteNoNo
Pre-indexed chunks + citationsYesOn-demand fetchSnippets only
Per-result freshness stampsEvery resultLast fetchSearch engine date
Site evaluation / scoringBuilt inNoNo

Best-effort comparison as of April 2026. Verify competitor tiers in their docs before making a buying decision.

The thing no single MCP can do

Search every indexed site at once.

A vendor MCP can answer questions about itself. Lattis answers questions across everyone. That's the capability no single-site integration can replicate — and the reason agent builders reach for it.

  • "Find a CRM with free tier + webhooks + great API docs."
  • "Which auth libraries support Cloudflare D1 natively?"
  • "Who charges a flat fee instead of a percentage?"
cross-site query
$ lattis search \
  "auth libraries with cloudflare d1"
3 matches across 14,000 sites
 • better-auth.com  native adapter
 • lucia-auth.com  community adapter
 • clerk.com   not supported

One endpoint today. Every site your agent will ever touch.

No per-site integrations. No DOM parsing on the agent's clock. No token bills you can't explain.