AI Agent Smart Code Search: Stop Burning Tokens on grep+read

Semble hit #6 on Hacker News with a shocking stat: AI agents waste up to 98% of tokens using grep+read to find code. Here's the smarter approach — and a Telegram agent that does semantic code search instead.

This week, Semble landed on Hacker News with 313 points. It's an open-source code search library built for AI agents, and its core claim stopped me cold:

AI agents using grep+read waste ~98% of tokens on irrelevant code. Semble's approach retrieves exactly what's needed at 99% of the retrieval quality of a 137M-parameter transformer, using 200x less compute.

Think about what that means. Every time your AI coding agent searches for a function, a class, or an implementation detail, it's grepping the entire codebase and then reading full files to find what it needs. In a large repo — tens or hundreds of thousands of lines — that's thousands of wasted tokens per query. Across a day of agentic coding, it adds up to real money.

The Token Waste Problem

Here's the math. When an AI agent needs to find "how authentication is handled" in a codebase, the fallback is usually:

1. grep for keywords like "auth" or "login" — returns dozens of matches
2. Read each matching file to understand context — potentially thousands of lines
3. Parse all that noise to extract the relevant 3-5 functions

Steps 2 and 3 are where the waste lives. The agent pays for every line it reads, even lines that have nothing to do with the question. Semble's benchmark shows that on a typical repo, grep+read consumes roughly 15,000 tokens per query vs. Semble's ~250 — a 98% reduction.

Why Semantic Code Search Matters for AI Agents

The deeper insight here isn't just about tokens. It's about agent architecture. An agent that has to grep and read full files is an agent that spends most of its context window on irrelevant content. That leaves less room for actual reasoning, planning, and code generation.

Semantic code search changes the equation. Instead of matching keywords, it understands intent. Ask "how do we handle rate limiting?" and it returns the actual middleware function, not every file that mentions "limit." The agent gets exactly what it needs in one shot, with 98% less noise.

With OpenClaw, you can build a Smart Code Search Agent on Telegram that handles this for your team: paste a natural language query about any codebase, and the agent retrieves the exact code you need — no grep skills required.

The Prompt

Copy-paste this into your OpenClaw Telegram bot. Send any code search question (with or without a repo URL) and get back the exact code snippets you need.

How to Use It

1. Deploy OpenClaw on GetClawCloud
2. Paste the prompt above as your agent's system prompt
3. Send a natural language code search — the agent returns structured results

Set Up the MCP Server (One-Time)

Semble slots into any MCP-compatible agent in one command. Run this on your dev machine:

# For Claude Code
claude mcp add semble -s user -- uvx --from "semble[mcp]" semble

# For Cursor (in .cursor/mcp.json)
{
  "mcpServers": {
    "semble": {
      "command": "uvx",
      "args": ["--from", "semble[mcp]", "semble"]
    }
  }
}

# For Codex CLI (in ~/.codex/config.toml)
[mcp_servers.semble]
command = "uvx"
args = ["--from", "semble[mcp]", "semble"]

Once configured, your agent will automatically use Semble for code search instead of falling back to grep+read. No prompts, no configuration — it just works.

Why Token Efficiency Is a Feature, Not a Detail

The AI agent economy runs on tokens. Every query you make costs something — either in API fees (for hosted models) or in context window space (for reasoning depth). A token-efficient agent isn't just cheaper; it's smarter, because it can spend its limited context on actual problem-solving instead of noise.

Semble's 98% reduction isn't an optimization you apply later. It's a fundamental architecture choice: give your agent the ability to find code semantically, and it will consistently produce better results, faster, at lower cost.

And because this runs on Telegram via OpenClaw, your entire team gets access from the chat app they already use. No new tools, no onboarding, no monthly seat fees.