My Web Search Kagi

Web Search via agent-browser

Perform web searches using browser automation. Kagi is preferred if available, Bing is the fallback that works without authentication.

⚡ Quick Start

Pick the fastest path for your setup:

⚠️ Important: CAPTCHA / Login Requirement

Kagi search via browser automation is blocked by Cloudflare Turnstile CAPTCHA unless the browser session has authenticated Kagi cookies.

The API path avoids this entirely. This skill provides four viable paths:

Path 0: API Direct (Fastest — Recommended for Programmatic Use)

If you have a Kagi API token, use the official API directly. It's ~10-50x faster (~100-500ms) than browser automation and returns structured JSON.

Prerequisites

• Kagi API token from kagi.com/settings?p=api
• Kagi Search API is now in public preview — all subscribers get $5 free API credits

Quick search with curl

# Set your token
export KAGI_API_KEY="your_token_here"

# Search
curl -H "Authorization: Bot $KAGI_API_KEY" \
  "https://kagi.com/api/v1/search?q=rust+tokio+async+runtime"

Python (official kagiapi package)

pip install kagiapi

from kagiapi import KagiClient
import os

kagi = KagiClient(os.environ["KAGI_API_KEY"])
results = kagi.search("rust tokio async runtime", limit=10)

for result in results["data"]:
    if result['t'] == 0:  # t=0 is search result, t=1 is related searches
        print(f"{result['title']}\n{result['url']}\n")

MCP Server (official kagimcp)

For Claude Desktop, Codex CLI, Cline, or any MCP client:

# Install uv first (if not already installed)
curl -LsSf https://astral.sh/uv/install.sh | sh

# Add to Codex CLI
codex mcp add kagi --env KAGI_API_KEY=<YOUR_API_KEY> -- uvx kagimcp

# Or add to Claude Desktop config (~/Library/Application Support/Claude/claude_desktop_config.json)
{
  "mcpServers": {
    "kagi": {
      "command": "uvx",
      "args": ["kagimcp"],
      "env": {
        "KAGI_API_KEY": "your_token_here"
      }
    }
  }
}

Tools exposed: kagi_search_fetch (web/news/videos/podcasts/images), kagi_extract (page as markdown)

API vs Browser Automation

Path 1: Firefox / Zen / LibreWolf / Waterfox users (Recommended for Browser Automation)

If the user uses a Firefox-variant browser and is already logged into Kagi, cookies can be synced into agent-browser automatically.

One-time setup

# Sync Kagi cookies from Zen (or firefox, librewolf, waterfox, floorp)
python3 ~/.pi/agent/skills/my-web-search-kagi/scripts/sync-kagi-cookies.py --browser zen

The script:

1. Finds the browser's default profile
2. Extracts Kagi cookies from cookies.sqlite
3. Injects them into the running agent-browser session via CDP

Search after cookie sync

agent-browser open "https://kagi.com/search?q=rust+tokio+async+runtime" && agent-browser get text

Path 2: Chromium / Chrome / Arc / Brave / Edge users

If the user uses a Chromium-based browser and is logged into Kagi there, reuse that profile directly.

One-time setup

# List available Chrome profiles
agent-browser profiles

# Use the profile that has Kagi logged in (e.g. "Default", "Profile 2", etc.)
agent-browser --profile "Profile 2" open "https://kagi.com"
# If Kagi shows you as logged in, you're done. Otherwise log in once.

Search with persistent profile

Always use --profile so the login state is preserved:

agent-browser --profile "Profile 2" open "https://kagi.com/search?q=QUERY" && agent-browser --profile "Profile 2" get text

To avoid typing --profile every time, set the env var:

export AGENT_BROWSER_PROFILE="Profile 2"
agent-browser open "https://kagi.com/search?q=QUERY" && agent-browser get text

Or set it in ~/.agent-browser/config.json:

{"profile": "Profile 2"}

Path 3: No Kagi login available → Bing fallback

If neither path 1 nor 2 is viable, use Bing as the search engine. Bing does not block automation and requires no authentication:

agent-browser open "https://www.bing.com/search?q=QUERY" && agent-browser get text

How to Search (General)

One-shot search + extract

# Kagi (after cookies synced or profile set)
agent-browser open "https://kagi.com/search?q=URL_ENCODED_QUERY" && agent-browser get text

# Bing (no setup needed)
agent-browser open "https://www.bing.com/search?q=URL_ENCODED_QUERY" && agent-browser get text

Interactive search

agent-browser open "https://kagi.com"
agent-browser type "input[placeholder*='Search']" "your query"
agent-browser press Enter
agent-browser wait 2000
agent-browser get text

Navigate to a specific result

agent-browser open "URL_FROM_RESULTS" && agent-browser get text

Best Practices

1. Use API first — if you have a token, it's 10-50x faster and more reliable
2. Try Kagi browser automation — better results, ad-free, respects user's subscription
3. Sync cookies for Firefox-variant users via the included script
4. Reuse Chrome profile for Chromium users via --profile
5. Fallback to Bing if Kagi auth is unavailable
6. Encode queries: Replace spaces with +, special chars with percent-encoding
7. Use get text: Clean innerText, no HTML clutter
8. Use snapshot: When you need to interact with specific elements
9. Chain searches: Search broad → extract URL → deep-dive

Limitations

• API requires token: Must get API key from Kagi settings
• Browser Kagi requires auth: Cloudflare Turnstile blocks unauthenticated automation
• Browser speed: ~2-5s per page (slower than API's ~100-500ms)
• Browser format: Raw text, not structured JSON
• Browser fragility: May break if search UI changes
• Cookie expiry: If Kagi sessions expire, re-run the sync script or re-login

Alternative: Kagi Search API (Now in Public Preview)

For high-volume or structured programmatic search, the Kagi Search API is now publicly available (no invite needed):

• Cost: $25 per 1,000 queries (2.5¢ each)
• Free credits: $5 added to all subscriber accounts (May 2026)
• Status: Public preview — ready for production use
• Docs: https://kagi.com/api/docs
• Python package: pip install kagiapi (github.com/kagisearch/kagiapi)
• MCP server: uvx kagimcp (github.com/kagisearch/kagimcp)
• Use when: Agent needs 10+ searches/hour or structured JSON results

See Path 0 above for setup instructions.

Troubleshooting

Common Agent Mistakes

❌ Wrong: Searching for information already in training data alone ✅ Right: Use search only for current/real-time info or specific facts not in training data

❌ Wrong: Using browser automation when you have an API token ✅ Right: Use the API (Path 0) for speed and reliability; save browser automation for when you don't have a token

❌ Wrong: Using Kagi browser automation without checking cookies first ✅ Right: Verify cookie sync worked before relying on Kagi results; fall back to Bing if needed

❌ Wrong: Falling back to Bing when Kagi API or browser automation works fine ✅ Right: Only use Bing fallback when Kagi is unavailable or unauthenticated

❌ Wrong: Using web search for tasks better suited to API calls ✅ Right: Use httpie/curl for known endpoints; web search is for discovery, not repeated API work

Related Skills & Tools

• @skills/my-crawl4ai — For crawling entire sites, not single search queries
• @skills/my-tech-stack — For tool recommendations (httpie vs curl, etc.)
• @skills/cua-computer-use — For GUI interactions when browser automation isn't sufficient
• Official Kagi MCP Server — uvx kagimcp for Claude Desktop, Codex CLI, Cline (github.com/kagisearch/kagimcp)
• Official Kagi Python Package — pip install kagiapi (github.com/kagisearch/kagiapi)