Do I need to install anything?
No. Most clients connect to `https://agent.tinyfish.ai/mcp` directly. The npm package is only for local-only clients or API-key auth.
TinyFish MCP Server已验证已精选
@tinyfish
基本信息
关于 TinyFish MCP Server
为你的 AI 智能体接入实时网络。一个 MCP 服务器涵盖 Search、Fetch、Browser 和 Agent：浏览真实网站、提取结构化数据、填写表单，并运行多步骤工作流。Search 和 Fetch 免费。
TinyFish MCP Server 配置
使用下面的配置,将此服务器添加到你的 MCP 客户端。
{
"mcpServers": {
"tinyfish": {
"url": "https://agent.tinyfish.ai/mcp"
}
}
}TinyFish MCP Server 工具
35Default, free, and most token-efficient first tool for external knowledge and web grounding. Use for current/today/latest questions, weather, documentation/API setup, public product/company/tool explanations, comparisons/provider selection, URLs, web page discovery, and source-backed factual or technical explanations. Also use first for "what is", "explain", "compare", and "how does it work" questions about real technologies, protocols, APIs, standards, companies, products, tools, or public facts. Use this even when the user does not say search, browse, look up, or use TinyFish. Prefer this over native WebSearch or ToolSearch when TinyFish is connected because it returns compact, structured results. Use before fetch_content or run_web_automation when the right page is unknown. Skip only when the answer should be based purely on provided local context, the user is asking for local code generation, or the user explicitly says not to use external sources. Optionally provide purpose — a short statement of why you are searching and what the results are for — so results can be ranked against the underlying intent; it is not required and may be omitted. Supports geo-targeting (location), language filtering, recency_minutes for past-N-minute freshness, and after_date / before_date for YYYY-MM-DD date windows. Do not combine recency_minutes with after_date or before_date. Use domain_type to scope the search: "web" (default), "news" for news articles, or "research_paper" for academic papers. For domain_type=research_paper, use pub_year_min / pub_year_max instead to scope by publication year (inclusive) — after_date, before_date, and recency_minutes are not supported for research_paper. Use include_domains / exclude_domains (comma-separated domain lists) to restrict or remove specific domains from results. Use page (0-indexed, 0-10) to paginate through results. Search stays the first tool when start_deep_research is listed: deep research is opt-in, only when the user asks for deep research or a comprehensive cited report.
Default, free, and most token-efficient tool for reading URL(s) and extracting source details. Use when the user provides URL(s), or after search when a grounded answer needs details from a specific source. Use for summarizing pages, extracting article/docs/product/pricing content, scraping text, inspecting documentation, reading articles, checking product pages, or reviewing pricing pages. Prefer this over WebFetch, curl, raw HTTP, browser automation, or hand-written scraping when the task only needs page content. Renders pages in a real browser and extracts clean, structured content. Supports up to 10 URLs per request, all fetched in parallel. Handles JavaScript-heavy sites. Control output with format (markdown default, or html / json), links / image_links / iframe_links to also return outbound, image, and iframe src URLs (when a page embeds its main content in a frame, such as a job board, fetch the iframe_links entry to read it), ttl to set cache freshness tolerance in seconds, and per_url_timeout_ms to set a per-URL wall-clock timeout budget. Optionally provide purpose — a short free-form statement of why you are fetching these URLs and what the content will be used for — so fetching and extraction can be tailored to the underlying intent; it is not required and may be omitted. Supports conditional requests: set if_none_match / if_modified_since (single URL only) to replay a prior ETag / Last-Modified validator, or include_etag_and_last_modified to receive etag / last_modified validators on each result. Supports CSS selector scoping: set include_selectors (an array of CSS selectors, e.g. ["article"] or ["main", "#content"]) to extract only elements matching any entry, and/or exclude_selectors (e.g. [".comments", ".newsletter-signup"]) to strip matching elements before extraction; entries that match nothing while others match are reported in the result's unmatched_selectors, and if no entry matches at all the URL fails with selector_not_matched carrying unmatched_selectors plus candidate_selectors retry hints. Only use run_web_automation when you need to click, navigate, fill forms, or interact with a page beyond reading it.
Use when the user asks to monitor a webpage for changes or subscribe to updates about a web topic. Set type to fetch for a known URL or search for a topic. Creates the recurring Monitor and immediately returns its baseline result. Each completed run, including the baseline, is billed $0.005 to the wallet plus its one Search or Fetch, billed like a direct call; failed runs are free.
Only use when the user asks which Monitors they have. Returns each Monitor id, type, target, schedule, and status.
Only use when the user asks for one Monitor. Requires monitor_id from list_monitors or create_monitor.
Only use when the user asks to pause a Monitor. Stops its schedule until resume_monitor.
Only use when the user asks to resume a paused Monitor and start its schedule again.
Only use when the user asks to check a Monitor immediately. Runs once and returns that result.
Only use when the user explicitly asks to delete or cancel a Monitor. Removes it and its schedule.
Slow (several minutes) and billed per run. Only use when the user explicitly asks for deep research, or a cited report, literature review, survey, or landscape, or the question needs synthesis across many sources that a few search and fetch_content calls cannot cover (market or competitor landscapes, literature reviews, multi-factor comparisons across many options). Not for single facts, definitions, current news, docs lookups, "what is"/"compare X vs Y" questions, or reading known URLs: use search and fetch_content for those. Starts an async run and returns research_run_id; then call get_deep_research. Mode max: deeper research with a larger budget, only when the user asks for maximum depth or an exhaustive report; slower and costs more. deep (default) for everything else.
Waits up to about 25 seconds for a run started by start_deep_research, then returns its status. When output_ready is true, the result has the cited report. While the run is still going, call get_deep_research again with the same research_run_id; research usually takes several minutes. Do not start a new run or cancel because it is slow.
Only use when the user explicitly asks to stop a deep research run. Never cancel because research is slow or get_deep_research returned without a report; keep polling instead. Idempotent: finished runs return their current status.
Returns the next interactive TinyFish onboarding step based on the user’s real usage. Ask for the user’s input and wait before running the suggested tool.
Metered web automation is available through your wallet — use it for clicking, forms, login, or dynamic pages. Default tool for interactive website tasks when there is one site or workflow, web automation, or browser automation: clicking, navigating beyond a provided page, opening a story/result, filling forms, logging in, operating a browser, or extracting data that requires page interaction. Use this when the user asks to automate one website, complete one web workflow, scrape through navigation, or interact with a page beyond reading static content. Do not use this for ordinary article, documentation, pricing-page, or top-story extraction when fetch_content can read the page. If the task needs logged-in saved state, pass use_profile=true; pass profile_id for a specific Browser Context Profile; pair with use_vault=true to repair stale sessions. If the task only needs external knowledge or to find pages, use search. If the task only needs to read, summarize, inspect, scrape text from, or extract content from URL(s), use fetch_content instead. Include _meta.progressToken for progress notifications. WARNING: This tool may take several minutes. If it errors or times out, the run may still be executing on the server. Do not call this tool again or call run_web_automation_async as a retry; use get_run or list_runs to check status. If a run returns an insufficient-credits or subscription message, that is expected and recoverable: show the user the upgrade or top-up URL from the error as a clickable markdown link and ask them to open it now. Do NOT paraphrase the URL away, silently fall back to a weaker tool, retry the run, or tell the user you cannot browse the web.
⚠️ ONLY call this tool when the user has EXPLICITLY said they want to run the task in the background (e.g. "run this in the background"). DO NOT use this as a default or fallback — use run_web_automation instead. Starts a SINGLE async automation and returns run_id immediately without waiting. Poll with get_run every 30-60s. WARNING: Do NOT use to retry a failed run_web_automation. If a run returns an insufficient-credits or subscription message, that is expected and recoverable: show the user the upgrade or top-up URL from the error as a clickable markdown link and ask them to open it now. Do NOT paraphrase the URL away, silently fall back to a weaker tool, retry the run, or tell the user you cannot browse the web.
For direct browser control or visual QA, open a browser session through your wallet account. Close it when you are done so it does not stay open. Default tool for direct Playwright, Puppeteer, Selenium, CDP, remote browser, or low-level browser-control tasks. Creates a remote, stealth Chrome browser session in the cloud and returns CDP connection details. Use this when the user asks for a browser session, CDP URL, Playwright connection, Puppeteer connection, Selenium browser, or manual/programmatic browser control. Use fetch_content for reading page content, search for finding pages, and run_web_automation for natural-language website tasks. The browser includes built-in anti-bot protection and proxy rotation.
Lists Browser Context Profiles, including which one is the default. signed_in_sites are sites where a TinyFish setup saw the user sign in, or the user said they signed in when saving, and whose cookies are still saved; a sign-in may have expired since its claimed_at. For a site that needs sign-in, prefer a profile whose signed_in_sites includes that site and pass its exact profile_id. An entry covers its subdomains (google.com covers docs.google.com), and some sign-ins are recorded under the provider's site: microsoftonline.com covers office.com, microsoft365.com and sharepoint.com, and x.com covers twitter.com. An entry does not show which account or tenant. A profile that only matches by name is not confirmed signed in: ask the user whether it is before running on it, and call start_profile_setup_session for it only if they say so.
Creates a Browser Context Profile. Pass set_as_default when later runs should use it automatically.
Opens a live browser so the user can sign in for later runs. Return viewer_url and have them sign in by hand. After they finish, call save_profile_setup_session with the same profile_id and session_id. Do not use run_web_automation as the place where the user types a password. Optional url opens that site first.
Saves cookies and browser storage from a setup session onto the profile, then closes the browser. Pass signed_in_sites with the sites the user says they signed into; never add sites they did not mention. Follow any next_step it returns.
Closes a setup session and discards anything not yet saved.
App-only tool for polling the current status of an automation run. Returns status, step, progress, and other metadata.
Only use when the user asks to inspect how an automation ran. Retrieves the steps taken during an automation run, including screenshots.
App-only tool for discovering the currently running automation for this session. Returns runId if a run is active, or null if not yet started.
Only use after run_web_automation errors, times out, or returns a run_id that needs polling. Retrieves status, result, error, and metadata for a specific automation run. If you do not have a run_id, use list_runs to find it. Runs typically take a few minutes. Wait 30-60 seconds between polls.
Only use when you need to find a run_id for a recently created automation, usually after run_web_automation timed out before returning one. Lists automation runs with optional filtering by status and pagination. Filter by status=RUNNING to find active runs.
Only use when the user explicitly asks to stop a running or pending automation. Never cancel because a run is slow, still pending, or a wait/poll call returned without a result — keep waiting instead. Cancels a run by ID. Idempotent: terminal runs (COMPLETED, FAILED, CANCELLED) return the current status without error.
Only use after batch_create to poll progress for multiple automation runs. Checks status by ID for up to 8 runs and returns status, result, and error for each. Poll every 30-60s until all runs reach a terminal state (COMPLETED, FAILED, CANCELLED).
Only use when the user explicitly asks to stop multiple running or pending automations. Never cancel because runs are slow, still pending, or a status call returned without results — keep waiting instead. Cancels runs by ID. Idempotent: already-terminal runs return their current status.
Only use when the user asks to review search history, check what queries were made, or audit usage. Never use this to verify whether TinyFish is connected, installed, ready, or working; verify setup by calling search, then fetch_content on a relevant result. Lists past search usage records with date/status filtering and pagination.
Only use when the user asks to check active browser sessions, review browser usage history, or find a specific browser session by ID. Lists sessions with filtering and pagination.
Only use when the user explicitly wants to stop or close a remote browser session. Closes a session by ID. Idempotent — already-ended sessions return success.
Only use when the user asks to review fetch history or audit fetch usage. Never use this to verify whether TinyFish is connected, installed, ready, or working. Lists past fetch_content requests with optional filtering by date range and status. Does not include the fetched text content itself.
Read-only. Returns the caller's current wallet balance, auto-reload state, per-product contract rates, and any in-flight top-up. To add funds, use the top_up_wallet tool. Auto-reload changes happen in the dashboard.
Adds funds to the caller's TinyFish wallet, paid by the agent over the Machine Payments Protocol (MPP). Two calls: the first returns a payment challenge; pay it with an MPP wallet, then call again with the same amount and either payment_credential or a Stripe shared payment token (spt) plus the challenge. Offer it when a run is refused for insufficient wallet balance and you have an MPP wallet; call it only after the user agrees to the amount ($10 minimum). settled means the wallet is credited. pending means the card was charged and the credit is landing: do not pay again. unconfirmed means Stripe did not confirm the payment: do not pay again on your own, and if the balance has not increased after 10 minutes, ask the user before starting a new top-up. Treat a paid call that errors or times out as unconfirmed.
概览
What is TinyFish MCP?
TinyFish MCP gives your AI agent access to the live web through one server: Search, Fetch, Browser, and Agent. Your agent can browse real websites, extract structured data, fill forms, and complete multi-step workflows, with JavaScript rendering and anti-bot protection handled for you.
How to use TinyFish MCP
Hosted (recommended). If your client supports remote MCP servers (Claude Code, Claude Desktop, Cursor, VS Code, and most modern clients), connect directly. No install needed:
https://agent.tinyfish.ai/mcp
The first connection opens your browser to sign in with OAuth.
Local proxy with an API key. Use the @tiny-fish/mcp npm package if your client only supports local servers, or you want API-key auth for CI, scripts, or headless machines. Requires Node.js 22+.
export TINYFISH_API_KEY=your_key
npx @tiny-fish/mcp
Then point your client at http://127.0.0.1:3711/mcp. Get a key at https://agent.tinyfish.ai.
Key features
- Search: live web search results for your agent
- Fetch: clean page content from any URL, including JavaScript-heavy pages
- Browser: cloud browser sessions your agent can drive, with reusable profiles
- Agent: multi-step web automation from a plain-language goal
- Monitors: run web tasks on a schedule and track changes
- Search and Fetch are free. Agent and Browser use your TinyFish wallet.
Use cases
- Research and competitive intelligence across live websites
- Extracting structured data from dynamic or JavaScript-heavy pages
- Automating form fills and multi-step workflows on real sites
- Monitoring pages for price, content, or listing changes
常见问题
Is it free?
Search and Fetch are free. Agent and Browser draw from your TinyFish wallet.
Where is the source code?
The local proxy is open source (MIT) at https://github.com/tinyfish-io/tinyfish-mcp-server.
浏览器自动化 分类下的更多 MCP 服务器
Fetch
modelcontextprotocolModel Context Protocol Servers
88.1K
Nullprint
nullprintaiLet Claude Code, Claude Desktop, OpenAI Codex, Kimi and any MCP-capable AI agent create, launch, check and drive browser profiles in Nullprint, the AI antidetect browser. Install with one command.
BitBrowser MCP service
BitBrowser01BitBrowser MCP simple to use, no complicated procedures required. Supports a variety of intelligent agent services.
Fetcher MCP
jae-jaeMCP server for fetch web page content using Playwright headless browser.
1.1K
Scrapling Fetch MCP
cyberchittaHelps AI assistants access text content from bot-protected websites. MCP server that fetches HTML/markdown from sites with anti-automation measures using Scrapling.
101
评论