component-search

# Component Search

Search 21st.dev's registry of production-ready React components. Returns ranked results with code, previews, and install instructions.

```bash
/ork:component-search animated pricing table
/ork:component-search sidebar with collapsible sections
/ork:component-search dark mode toggle switch
```

## How It Works

```
Query: "animated pricing table with monthly/annual toggle"
  │
  ▼
┌──────────────────────────────┐
│ 21st.dev Magic MCP           │  Search the 21st.dev developer registry
│ @21st-dev/magic              │  Filter: React, Tailwind, shadcn
└──────────┬───────────────────┘
           │
           ▼
┌──────────────────────────────┐
│ Results (ranked by relevance)│
│                              │
│ 1. PricingToggle (98% match) │  ★ 2.3K views · shadcn/ui
│ 2. PricingCards (87% match)  │  ★ 1.8K views · Radix
│ 3. AnimatedPricing (82%)     │  ★ 950 views · Motion
└──────────────────────────────┘
```

## Step 0: Parse Query

```python
QUERY = ""  # Component description

# 1. Create main task IMMEDIATELY
TaskCreate(subject="Component search: {QUERY}", description="Search 21st.dev registry", activeForm="Searching for {QUERY}")

# 2. Create subtasks for each phase
TaskCreate(subject="Parse query and detect project context", activeForm="Detecting project context")  # id=2
TaskCreate(subject="Search component registry", activeForm="Searching registry")                      # id=3
TaskCreate(subject="Present and deliver results", activeForm="Presenting results")                    # id=4

# 3. Set dependencies for sequential phases
TaskUpdate(taskId="3", addBlockedBy=["2"])  # Search needs project context first
TaskUpdate(taskId="4", addBlockedBy=["3"])  # Results need search done

# 4. Update status as you progress
TaskUpdate(taskId="2", status="in_progress")  # When starting
TaskUpdate(taskId="2", status="completed")    # When done — repeat for each subtask

# Detect project context for framework filtering
Glob("**/package.json")
# Read to determine: React version, Tailwind, shadcn/ui, styling approach

# Detect shadcn/ui style for result ranking
Glob("**/components.json")
# Read → "style" field (e.g., "radix-luma", "base-nova")
# Used to prefer components matching the project's visual language
```

## Step 1: Search Registry

### Browse first, and let the USER pick

**The reason is choice, not cost.** Search returns a preview image, a video, the
author and the `installCommand` — enough for a human to recognise the right
component at a glance. The model cannot. Guessing which of eight "command palette"
results the user meant is the failure mode this flow exists to prevent, and it
stays a failure mode on every pricing tier.

**If 21st-dev-magic MCP is available (the real path):**

```python
# 1. BROWSE. Prefer search_picker: same params and results as search, but it
#    renders an inline picker so the USER chooses. Do not pre-filter to one
#    result on their behalf — show the options.
mcp__21st-dev-magic__search_picker(query="{QUERY}", type="component", limit=8)

# 2. PICK — the user selects. Stop here and wait.

# 3. RETRIEVE — the chosen component.
mcp__21st-dev-magic__get_component(id=<demo id from the chosen result>)
```

- If the user only needs to know *whether* something exists, or wants the install
  command, **stop after step 1** — search already answered it.
- `search` returns an `installCommand` containing `?api_key=$API_KEY_21ST`. That
  is a **different env var** from the MCP's `TWENTY_FIRST_API_KEY`; if
  `npx shadcn add` fails auth, that variable is missing from the environment.

### Tier — check it, do not assume it

`get_component` is the only metered tool; `search`, `search_picker`, `search_logo`,
`get_theme` and every list/metadata tool are unmetered on all tiers.

**Call `get_usage` rather than assuming a cap.** It returns `tier` plus
`freeRetrievalsRemaining` (both `null` when unlimited).

| tier | `get_component` |
|---|---|
| `free` | **2 / day** — treat each retrieval as the day's budget: never speculative, never in a loop, never "to compare" |
| `paid` | unlimited — comparing two candidates is fine; the picker step still applies |

Written 2026-08-16 against a free account and corrected 2026-08-17 when the
account went paid: the original text argued the flow from the 2/day cap, which
made it read as obsolete the moment the cap lifted. The cap is a constraint;
letting the user choose is the design.

**If 21st-dev-magic is NOT available (fallback):**
```python
# Genuine fallback ONLY — scraping returns no ids, no install commands, and
# no structured metadata. If the MCP is configured, you should never be here.
WebSearch("site:21st.dev {QUERY} React component")
# Or browse the registry
WebFetch("https://21st.dev", "Search for: {QUERY}")
```

**Alternative generation path — v0.app MCP (2026):**
When the registry doesn't have a matching component, fall back to AI
generation via the v0.app MCP server rather than WebFetch scraping:

```python
# If @vercel/v0-mcp is available
# v0.app generates from the same query and can pin to shadcn style
# (e.g., luma / nova / lyra) via `shadcn apply` after download.
```

This is a **generation** path, not a registry search — results will not
have view counts or stars. Prefer registry search when possible so you
get battle-tested components; use v0.app only when the registry misses.

## Step 2: Present Results

Show top 3 matches with:
- Component name and description
- Match relevance score
- Popularity (views/bookmarks)
- Framework compatibility
- Preview (if available)
- Install command

```python
AskUserQuestion(questions=[{
  "question": "Which component to use?",
  "header": "Component",
  "options": [
    {"label": "{name_1} (Recommended)", "description": "{desc_1} — {views_1} views"},
    {"label": "{name_2}", "description": "{desc_2} — {views_2} views"},
    {"label": "{name_3}", "description": "{desc_3} — {views_3} views"},
    {"label": "None — generate from scratch", "description": "Build a cust