Use Tinyfish

TinyFish CLI

The complete web toolkit — four tools, one CLI. Start with the lightest tool that can do the job and escalate only when needed.

Pre-flight Check (REQUIRED)

Before making any TinyFish call, always run BOTH checks:

1. CLI installed?

which tinyfish && tinyfish --version || echo "TINYFISH_CLI_NOT_INSTALLED"

If not installed, stop and tell the user:

Install the TinyFish CLI: npm install -g @tiny-fish/cli

2. Authenticated?

tinyfish auth status

If not authenticated, stop and tell the user:

You need a TinyFish API key. Get one at: https://agent.tinyfish.ai/api-keys

Then authenticate:

tinyfish auth login

Do NOT proceed until both checks pass.

Picking the Right Tool

search  →  fetch  →  agent  →  browser
lightest                        heaviest

Common Patterns

Research: search → fetch Search for a topic, then fetch the best results to read their full content.

# 1. Find URLs
tinyfish search query "best React state management libraries 2026"

# 2. Read the top results
tinyfish fetch content get --format markdown "https://result1.com" "https://result2.com"

Deep extraction: search → agent Search to find the right site, then use agent to interact with it and extract structured data.

# 1. Find the site
tinyfish search query "Nike running shoes official store"

# 2. Automate extraction on it
tinyfish agent run --url "https://nike.com/running" \
  "Extract all running shoes as JSON: [{\"name\": str, \"price\": str, \"colors\": [str]}]"

Escalation: fetch → agent Try fetch first. If the page is dynamic/JS-heavy and fetch returns empty or incomplete content, escalate to agent.

Full control: agent → browser If agent can't handle a complex multi-step workflow, spin up a raw browser session and automate it yourself via CDP.

Commands

tinyfish search query

Web search. Returns ranked results with titles, URLs, and snippets.

tinyfish search query "<query>" [--location <hint>] [--language <hint>] [--pretty]

• Returns 10 results by default
• Use --location and --language for geo-targeted results
• Default output is JSON; --pretty for human-readable

tinyfish search query "best pho in Ho Chi Minh City" --location "Vietnam" --language "en"

tinyfish fetch content get

Fetch clean, extracted content from one or more URLs. Strips ads, nav, boilerplate — returns just the content.

tinyfish fetch content get <urls...> [--format markdown|html|json] [--links] [--image-links] [--pretty]

• Accepts multiple URLs in a single call — they are fetched in parallel server-side
• --format markdown (default) — clean readable text
• --format json — structured document tree
• --links — include all extracted links from the page
• --image-links — include extracted image URLs
• Response includes: url, final_url, title, language, author, published_date, text, latency_ms

# Fetch one page as markdown
tinyfish fetch content get --format markdown "https://example.com/article"

# Fetch multiple pages with links
tinyfish fetch content get --links "https://site-a.com" "https://site-b.com" "https://site-c.com"

tinyfish agent run

Run a browser automation using a natural language goal. The agent opens a real browser, navigates, clicks, fills forms, and extracts data.

tinyfish agent run --url <url> "<goal>" [--sync] [--async] [--pretty]

Output: Default streams data: {...} SSE lines. The final result is the event where type == "COMPLETE" and status == "COMPLETED" — the extracted data is in the resultJson field. Read the raw output directly; no script-side parsing is needed.

Always specify the JSON structure you want in the goal:

tinyfish agent run --url "https://example.com/products" \
  "Extract all products as JSON array: [{\"name\": str, \"price\": str, \"url\": str}]"

tinyfish agent run --url "https://example.com/search" \
  "Search for 'wireless headphones', filter under $50, extract top 5 as JSON: [{\"name\": str, \"price\": str, \"rating\": str}]"

Parallel extraction — when hitting multiple independent sites, make separate calls. Do NOT combine into one goal.

Good — parallel calls (run simultaneously):

tinyfish agent run --url "https://pizzahut.com" \
  "Extract pizza prices as JSON: [{\"name\": str, \"price\": str}]"

tinyfish agent run --url "https://dominos.com" \
  "Extract pizza prices as JSON: [{\"name\": str, \"price\": str}]"

Bad — single combined call:

# Don't do this — less reliable and slower
tinyfish agent run --url "https://pizzahut.com" \
  "Extract prices from Pizza Hut and also go to Dominos..."

Managing runs:

tinyfish agent run list [--status PENDING|RUNNING|COMPLETED|FAILED|CANCELLED] [--limit N]
tinyfish agent run get <run_id>
tinyfish agent run cancel <run_id>

Batch operations — submit many runs from a CSV file (url,goal columns):

tinyfish agent batch run --input runs.csv
tinyfish agent batch list
tinyfish agent batch get <batch_id>
tinyfish agent batch cancel <batch_id>

tinyfish browser session create

Spin up a remote browser instance. Returns a CDP WebSocket URL for programmatic control.

tinyfish browser session create [--url <url>] [--pretty]

• --url optionally navigates to a page after creation
• Returns session_id, cdp_url (WebSocket), and base_url
• Use the cdp_url with Playwright, Puppeteer, or any CDP client

tinyfish browser session create --url "https://example.com"
# Returns: { session_id, cdp_url: "wss://...", base_url: "https://..." }

Getting Unstuck (consent walls & CAPTCHAs)

Live sites throw up blockers that stall agents. Baseline behavior for agent and browser runs:

• Dismiss consent/cookie/age walls automatically — make it the first step of every goal (Google "Before you continue", GDPR banners, region and "open in app" modals). Prefer "Reject all" where offered.
• On a CAPTCHA or bot check (reCAPTCHA, hCaptcha, Cloudflare "Checking your browser", press-and-hold): let the real browser clear invisible checks, retry at most once, then hand off to a human — return a BLOCKED status with the URL, or open a viewable browser session for the user to solve it, then resume. Never loop on a CAPTCHA; that is the classic stuck state. Logins/2FA/OTP are always user-provided.
• Every run ends in a defined state — DONE / BLOCKED / FAILED with a per-step timeout, so nothing hangs.

The full playbook (detection signals, escalation ladder, goal templates) lives in the browser-tasks skill at references/unblocking.md. Automate only what the user could legitimately do themselves.

General Notes

• Match the user's language: Respond in whatever language the user writes in.
• All commands support --pretty for human-readable output. Default is JSON.
• Use --debug on the root command or set TINYFISH_DEBUG=1 to log HTTP requests to stderr.

$ARGUMENTS