| name | kagi-usage |
| description | Search the web, extract or summarize a page, translate it, or check the news with the kagi CLI instead of WebSearch or WebFetch. Load this first for any Kagi task: it covers search and page content, and routes Assistant, automation, and account work to the right skill. |
| allowed-tools | Bash(kagi:*) Skill(kagi:*) |
Kagi CLI
Search the web and read pages with kagi. Route every other Kagi task to the
skill named below.
Route the task
| User intent | Where |
|---|
| Find, compare, or verify information from the web | Search the web (below) |
| Read, summarize, question, or translate a known page | Work with one page (below) |
| Use paid AI for reasoning, coding, research, or files | kagi-ai |
| Continue a conversation or manage Assistant state | kagi-assistant |
| Run searches in bulk or monitor changes over time | kagi-monitoring |
| Configure credentials or account search behavior | kagi-account-config |
Load the skill when it appears in the available skills listing. Otherwise run
kagi skills get <name> for the same guidance.
kagi skills list
kagi skills get kagi-assistant
Search the web
Discover sources, narrow the question, and return evidence that matches the
depth the user asked for.
| Need | Command |
|---|
| Direct answer with references | kagi quick |
| Ranked web results | kagi search |
| Search and summarize top pages | kagi search --follow N |
| Current news stories | kagi news or kagi search --news |
| Independent personal sites | kagi smallweb |
| Several independent queries | kagi batch |
quick, session-only search controls, and the subscriber search path require
session auth.
The current Search API uses KAGI_API_KEY. Public news and smallweb need no
credentials.
- Define the claim or question that needs evidence.
- Start with one broad search, or
quick for a bounded factual question.
- Split the topic into distinct queries when one query cannot cover it.
- Prefer primary sources and direct evidence.
- Use
--follow only when the user needs the top result pages summarized.
- Report source disagreement instead of flattening it.
kagi search "rust async cancellation" --format toon --limit 5
kagi quick "what is the current Kagi Search API endpoint?" --format markdown
Search controls
kagi search "query" --region us --format toon --limit 10
kagi search "query" --time month --order recency --format json
kagi search "query" --snap reddit --format toon
kagi search "query" --lens 2 --format toon
kagi search "query" --follow 3 --format markdown
Session-only controls include --lens, --time, --order, --verbatim,
personalization flags, and News vertical search. Use --local-cache only when
stale results are acceptable.
Query design
- Quote exact error messages or disputed phrases.
- Add official domains with
site: when a primary source is known.
- Use native-language query terms for regional or multilingual research.
- Separate discovery, counterevidence, and recency into different queries.
- Use
--region as a result bias, never as a language guarantee.
kagi batch \
"topic official documentation" \
"topic limitations" \
"topic independent review" \
--format toon --limit 5
News and discovery
kagi news --category tech --limit 10
kagi news --list-categories
kagi news --chaos
kagi search "open source ai" --news --format toon
kagi smallweb
Use news for the public Kagi News feed. Use search --news when the query must
filter the News vertical.
Work with one page
Work from the source the user supplied. Search the web only when the task needs
evidence that page cannot give.
| Outcome | Command | Credential |
|---|
| Full readable page markdown | kagi extract | KAGI_API_KEY |
| Summary of a URL or text | kagi summarize | session or legacy API token |
| Answer about one page | kagi ask-page | session token |
| Translation | kagi translate | session token |
Extract a page
Use extract when downstream work needs the article body rather than a
summary. Prefer extract over browser scraping for readable main content. It
requires the current KAGI_API_KEY, never the legacy KAGI_API_TOKEN.
kagi extract "https://example.com/article"
Summarize
Prefer subscriber mode when a session token is available.
kagi summarize --subscriber --url "https://example.com/article"
kagi summarize --subscriber --text "long text"
kagi summarize --subscriber \
--url "https://example.com/article" \
--summary-type keypoints
Use the public API mode only when the legacy API token is configured or the
task needs its engine controls.
kagi summarize --url "https://example.com/article"
kagi summarize --text "long text" --engine cecil
Never summarize content that must be quoted or checked precisely. Extract it
and inspect the relevant passage instead.
Ask one page
Use ask-page for a focused question whose evidence should come from one URL.
kagi ask-page \
"https://example.com/article" \
"what evidence supports the main claim?" \
--format markdown
When the question needs outside evidence or comparison, search the web instead.
Translate
kagi translate \
--text "Bonjour tout le monde" \
--target-language EN
Preserve code, URLs, product names, and other text that must stay untranslated.
State the target language explicitly.
Shared rules
- Run
kagi auth status before treating a command as unavailable.
- Choose the narrowest command that produces the requested outcome.
- Use
--format json for programmatic parsing.
- Use
--format toon for compact LLM context.
- Use
--format markdown for prose that a person will read.
- Run
kagi auth check when credentials exist but a request fails.
- Never print credential values.
kagi auth status
kagi auth check
Completion criteria
Web research is complete when:
- the answer addresses the exact question;
- important claims trace to source URLs;
- source quality and recency match the claim;
- disagreement and uncertainty appear in the output; and
- the output uses the requested format and depth.
Page work is complete when:
- the command matches the requested transformation;
- the output stays grounded in the supplied URL or text;
- quotes remain distinguishable from summaries;
- the output keeps the requested language and format; and
- you report a source-access failure instead of guessing past it.
