Modified
The model is the brain. The ecosystem is hands and eyes.
An has trained on enormous amounts of text and data. It can reason, write code, translate, summarize — essentially across any topic imaginable.
What it can't do is act in your reality — open the calendar, read today's news, click around in a browser. For that it needs tools. The ecosystem is the collection of those tools, and the five verbs to the right cover most of it.
01 — Automation · Wire up the apps
Automation platforms let you connect , databases and web services without writing everything from scratch. and are no-code options for fast . (see the Agents page) is open source with full code control.
Zapier
The world's largest automation platform. 6,000+ . The easiest way to connect AI to existing tools like Gmail, Slack and CRM systems. Glue layer, no code required.
02 — MCP · Plug in superpowers
is a standard protocol for tools. Instead of every AI app building its own integrations to Slack, GitHub and the filesystem, there is a shared interface. Plug in an MCP server — the agent gets the superpower instantly. Open standard, every LLM can implement it.
Playwright MCP
Browser control via MCP. Your can navigate pages, click buttons and extract data, just like a human. From .
Filesystem MCP
Gives local agents secure access to the filesystem via MCP. Read, write and search files without exposing the entire system. From Anthropic.
GitHub MCP
Repo operations via MCP: create issues, review pull requests, read code. Connects your agent directly to GitHub's API with a standardized interface.
Fetch MCP
HTTP requests as an MCP tool. The agent can fetch web pages and responses without an external browser. The simplest way to give an agent access to the web. From Anthropic.
03 — Search engines · Search the web
A model only knows what it was trained on. Search engines give the agent real-time access — facts from today, with sources. Most are based in the US. If you handle European user data: is EU-native and can run on your own infrastructure.
Tavily
The default for AI agents. Optimized for LLM responses with structured results and source citations. The easiest to integrate into agent pipelines.
Brave Search
Independent search index, not dependent on or Bing. Good for agents that need unfiltered search results without personalization.
Exa
optimized for pipelines. Returns full documents, not just snippets. Excellent when your agent needs deeper context.
Perplexity API
Search and summarization in a single API call. The agent gets an answer with sources, not raw search results that need to be processed.
04 — Scraping · Fetch data from sites
The difference from search: a scraper fetches the whole page and turns it into clean text the agent can read. Think policy documents, price lists, product data. Firecrawl is the most agent-friendly — -rendered pages and PDFs just work.
Firecrawl
Turns any URL into clean Markdown that LLMs can read. Handles JavaScript-rendered pages, auth walls and PDFs. The most agent-friendly in its class.
Apify
A marketplace of ready-made scrapers ("Actors") for Amazon, LinkedIn, TikTok and hundreds of other sites. Runs in the cloud — no servers of your own.
Crawlee
A crawling library in Node.js from Apify. and Cheerio integration, intelligent queue management and retry logic. For when you want to build your own scrapers with full control.
05 — Web agents · Click like a human
Half the web has no API. Web agents drive a real browser — clicking, filling forms, logging in. Browserbase gives the agent a cloud Chrome without you having to host it. Stagehand takes natural language: "click Log in" works without CSS selectors.
Browserbase
Cloud-hosted browser for agents. The API drives a real Chrome instance — perfect when you do not want to manage infrastructure yourself.
Stagehand
Built on top of Playwright. Lets an LLM drive the browser with natural language — "click Log in" works without CSS selectors.