---
layout: tap
site_name: tap
tap_name: filter
description: "Filter rows where field matches condition (gt, lt, eq, contains)"
intent: read
columns: []
args:
  - name: field
    type: string
    description: "Column name to filter on"
  - name: gt
    type: number
    description: "Greater than (numeric)"
  - name: lt
    type: number
    description: "Less than (numeric)"
  - name: eq
    type: string
    description: "Equals (exact match)"
  - name: contains
    type: string
    description: "Contains substring"
args_json: |
  {"field":{"type":"string","description":"Column name to filter on"},"gt":{"type":"number","description":"Greater than (numeric)"},"lt":{"type":"number","description":"Less than (numeric)"},"eq":{"type":"string","description":"Equals (exact match)"},"contains":{"type":"string","description":"Contains substring"}}
health_json: |
  {"min_rows":1,"non_empty":[]}
example_args: ""
source_url: https://github.com/LeonTing1010/tap-skills/blob/main/showcase/tap/filter.plan.json
license: MIT
---

<section>
  <h2>What it does</h2>
  <p>Filter rows where field matches condition (gt, lt, eq, contains)</p>
</section>

<section>
  <h2>Install Taprun once</h2>
  <p>Taprun ships as a single MCP server exposing a catalog of compiled taps. One-time setup on macOS / Linux:</p>
  <pre><code>brew install LeonTing1010/tap/taprun
tap mcp connect</code></pre>
  <p>Or drop this into your <code>claude_desktop_config.json</code> (works identically in Claude Code, Cursor, Cline, Windsurf — any MCP host):</p>
  <pre><code>{
  &quot;mcpServers&quot;: {
    &quot;tap&quot;: {
      &quot;command&quot;: &quot;tap&quot;,
      &quot;args&quot;: [&quot;mcp&quot;, &quot;start&quot;]
    }
  }
}</code></pre>
</section>

<section>
  <h2>Call <code>tap/filter</code></h2>
  <p>Terminal, once installed:</p>
  <pre><code>tap run tap/filter</code></pre>
  <p>From the MCP host — exact same compiled plan, deterministic replay, zero LLM tokens:</p>
  <pre><code>tap.run({ site: &quot;tap&quot;, name: &quot;filter&quot; })</code></pre>
</section>

<section>
  <h2>Why compile it once</h2>
  <p>This plan was forged once — the AI read <code>tap</code>, picked stable structural addresses (JSON-LD, ARIA, RSS, or declared API endpoints, in that priority order), and saved them to a <code>.plan.json</code>. Every replay since then has used zero LLM tokens. When <code>tap</code> ships a site change that breaks the extraction, <code>tap verify</code> surfaces it before your data goes stale — not after your pipeline silently writes garbage for a week.</p>
</section>

<section>
  <h2>Related tap taps</h2>
  <table><thead><tr><th>tap</th><th>description</th></tr></thead><tbody>
    <tr><td><a href="/taps/tap/table"><code>tap/table</code></a></td><td>Format rows as a human-readable table (outputs one row per line)</td></tr>
    <tr><td><a href="/taps/tap/sort"><code>tap/sort</code></a></td><td>Sort rows by field (numeric or alphabetic, asc or desc)</td></tr>
    <tr><td><a href="/taps/tap/pick"><code>tap/pick</code></a></td><td>Select specific columns from rows (projection)</td></tr>
    <tr><td><a href="/taps/tap/limit"><code>tap/limit</code></a></td><td>Take first N rows (head) or skip M then take N (offset+limit)</td></tr>
    <tr><td><a href="/taps/tap/dedupe"><code>tap/dedupe</code></a></td><td>Remove duplicate rows by field value</td></tr>
  </tbody></table>
</section>