Home › Agent Infrastructure

Agent Infrastructure for Claude

50 deterministic tools that make Claude agents faster, cheaper, and more reliable. One API key. $19/month flat. Measured 90% token cost reduction.

90%
Token cost reduction
50
Infrastructure tools
$19
Flat monthly price
<1ms
Cache hit latency

What is Agent Infrastructure?

When you build Claude-powered agents, the biggest cost driver is not the reasoning — it is the input. Agents re-read the same web pages, re-parse the same documents, re-run the same searches, and carry redundant context into every turn. Each redundant token costs money.

Agent infrastructure is the layer of deterministic, cacheable tools that sits between your Claude agent and the world. It shrinks Claude's input by handling the mechanical work — fetching, caching, compressing, slicing — so Claude only sees what is new and relevant.

KWJ is a hosted agent infrastructure API. You get 50 production-ready tools via one API key, without building or operating anything yourself.

The Token Cost Problem

A typical Claude agent session has three major token sinks:

  1. Redundant web content — fetching the same URL multiple times, or fetching full pages when you only need a CSS selector
  2. Verbose output — raw command output, full log files, and uncompressed API responses that fill the context window
  3. Re-derived answers — computing the same result (file outline, code slice, web extract) on every turn instead of caching it

KWJ eliminates all three categories with purpose-built tools that are measurably cheaper than letting Claude handle them inline.

Core Agent Infrastructure Tools

web

Web Cache

Fetch any URL and cache the response by content hash with configurable TTL. Repeated fetches of the same URL cost zero tokens.

compression

Output Compressor

Collapse repeated lines, elide middles, always rescue error/warn/fail lines. Turns 10,000-line build logs into 50-line digests.

code

Code Slicer

Extract a single function, class, or symbol from any Rust/Python/TS/Go file. Read 20 lines instead of 6,000.

cache

Answer Cache

Content-addressed result cache with fuzzy/semantic lookup via Jaccard shingle similarity. Never pay twice for the same expensive analysis.

docs

Document Parser

PDF, CSV, XLSX, DOCX extraction with table/search/convert/summary. Parse once, cache forever.

mcp

MCP Bridge

All 50 tools auto-register as MCP (Model Context Protocol) endpoints. One API key, one mcpServers entry in your Claude config.

memory

Memory Recall

TF-IDF semantic search over your agent's memory files. Pull only the relevant facts instead of loading everything into context.

vault

Secret Store

AES-256-GCM encrypted credential store with audit logging. Your agent can read API keys without storing them in prompts.

How It Works

1. Get an API key

Enter your email at kwj.ai/buy. Your API key is generated instantly. No credit card required for the 1-hour free trial.

2. Add to your Claude MCP config

{
  "mcpServers": {
    "kwj": {
      "command": "npx",
      "args": ["-y", "@kwj/mcp"],
      "env": { "KWJ_API_KEY": "kwj_your_key_here" }
    }
  }
}

3. Tools are available immediately

All 50 KWJ tools appear as MCP tool calls in your Claude environment. Claude can call web_read, cache_get, slice, digest, and the other 46 tools with no additional setup.

4. Measure the savings

KWJ includes a token meter tool (meter_report) that shows exactly how many tokens each tool saved per session. Typical savings: 60-90% on context-heavy agent workflows.

Integration Patterns

Web research agents

Call web_read instead of fetching URLs inline. Results are cached by URL hash with a configurable TTL (default 1 hour). A research agent that visits 20 URLs per run costs zero re-fetch tokens on the second run.

Codebase agents

Call repo_map to get a symbol outline of any directory, then slice to pull only the function you need. A 6,500-line main.rs becomes a 30-line symbol list + 50-line function slice — a 95% input reduction.

Build and CI agents

Pipe cargo build and cargo test output through digest. Failures surface immediately; passing tests are collapsed to a one-line summary. Context per build turn drops from 3,000 tokens to 80.

Document analysis agents

Call doc_parse to extract structured text from PDFs and spreadsheets. Results are cached by file hash. Parse a 200-page PDF once; subsequent calls return instantly from cache.

Comparison: With and Without KWJ

A typical Claude Code session on a large codebase:

At Anthropic Sonnet pricing, that is the difference between $7.80 and $0.78 per session.

Enterprise Agent Infrastructure

For teams building production Claude agents at scale, KWJ Enterprise adds:

Learn more about KWJ Enterprise →

Start Building Better Agents

Free 1-hour trial. $19/month after. Cancel anytime. Your API key is generated instantly.

Get Your API Key →