--- name: exa-agent-cli description: Use the exa-agent CLI for Exa search, contents, answer, context, similar, Websets, monitors, and agent workflows. --- # exa-agent CLI skill Generated by `cargo xtask generate-skill`; do not edit by hand. ## Guidance - Use capabilities first to discover command metadata. - Use capabilities (for example capabilities search) for a small command slice. - Use --dry-run --print-request before live mutations. - Search is not cursor-paginated: use --num-results and follow error.suggestedCommand when an invocation is rejected. - Search returns query-aware 800-char highlights by default; use --no-highlights for metadata only, or --text 1500 instead of --text full for capped triage text. - Search results are under `.data.results[]`; verify the live JSON path with `exa-agent search "rust async runtimes" --num-results 1 --json | jq '.data.results[] | {title,url}'`. - Filter search with `exa-agent search "AI infrastructure" --include-domain "exa.ai" --num-results 5 --json`. - SOURCE_NOT_AVAILABLE is not a zero-result success. Broaden and filter locally: `exa-agent search "AI infrastructure" --num-results 20 --json | jq '[.data.results[] | select(.url | test("^https?://([^/]+\\.)?exa\\.ai(/|$)"; "i"))]'`; cite the accessible publisher rather than treating a syndicator as the original source. - Contents accepts positional URLS or `--ids`: `exa-agent contents "https://exa.ai" "https://docs.exa.ai" --text 10000 --json`; text accepts bare, `full`, or numeric caps 1..10000. - --ndjson emits one object per result for list-shaped data and a final summary envelope; non-list commands fall back to compact JSON. - Contents/fetch success envelopes add outcome no_content (no failures and no returned rows), partial, or full (one results row per requested item with no failure evidence); missing statuses do not downgrade full coverage, and all-URL failures remain outcome partial with exit 10. - Empty contents error objects use upstream_reason_unavailable and suggest retrying or direct-fetching the quoted URL. - Set EXA_AGENT_NO_NETWORK to any value (including empty) to refuse live typed, raw, streaming, auth test/status, schema refresh --check, and doctor --online before credential resolution and transport; unset it to allow live calls, while dry-run and self-description remain available. - Do not pass managed auth headers; use EXA_API_KEY or auth login. - Errors are JSON on stderr with stable error.code values; run robot-docs errors for the full dictionary. ## Offline examples ```sh exa-agent capabilities --compact ``` ```sh exa-agent capabilities search --compact ``` ```sh exa-agent search "AI infrastructure news" --include-domain "exa.ai" --num-results 5 --dry-run --print-request --compact ``` ```sh exa-agent contents "https://exa.ai" "https://docs.exa.ai" --text 10000 --dry-run --print-request --compact ``` ```sh exa-agent team info --dry-run --print-request --compact ``` ## Footguns (bakeoff-learned, 2026-07-17) - There is NO `github` search category. Valid: company, people, research paper, news, personal site, financial report. For GitHub releases/changelogs, skip `--category` entirely. - `answer` finds pointers but does not dig page bodies (it will say so itself). For release notes, statutes, or any quotable body text, chain: `answer`/`search` to get the URL, then `contents "" --text`. - For statutory/official text, force authoritative domains (`--include-domain uscode.house.gov`, govinfo.gov, efile.fara.gov, agency sites) — the index happily serves clean-looking aggregator mirrors otherwise.