Research
2 tools to search the web and read pages as clean text. Other packs build on them, for example to research a company before you add it to your pipeline.
Research loads at startup by default. It works with no key.
Tools
| Tool | Level | What it does |
|---|---|---|
web_search | READ | Search the web. Returns titles, URLs and snippets |
web_fetch | MEDIUM | Fetch a public page and return its readable text |
web_search
| Argument | Notes |
|---|---|
query | What to search for |
max_results | 1-20, default 5 |
domains | Only search these sites, for example ["github.com"] |
Search uses Tavily when TAVILY_API_KEY is set. Tavily has a free tier and
gives better results. Without a key, Causly falls back to DuckDuckGo, which needs no account but
may be rate limited.
causly-server keys set TAVILY_API_KEY
web_fetch
Fetches a public http or https page and returns its title, description and text.
| Argument | Notes |
|---|---|
url | The page to read |
max_chars | Characters of text to return, 500-30,000, default 8,000 |
include_links | Also return up to 50 links from the page |
Private and local addresses are blocked, so web_fetch cannot be used to reach services on your
own network.
Untrusted content
Web pages are written by other people, and some try to give instructions to AI agents. Results
from both tools are marked as untrusted, and your agent is told never to follow instructions found
in them. Keep this in mind when you let an agent act on what it reads, and keep confirm
decisions with you. See Security.
Examples
"Search for open-source alternatives to Calendly and summarise the top three."
"Read acme.com/about and tell me what they do."