Web Tools

Web search and validated web-page fetching

What it does: Gives agents internet access — web_search queries search engines, web_fetch reads page content. Both tools include security protections against accessing internal networks.

Who it’s for: Pretty much every agent. Web tools are the most commonly used agent capabilities — they let your agent look up current information, read documentation, check prices, verify facts, and stay informed about topics that go beyond its training data.

Both tools are enabled by default in the full tool policy profile and ship in the cron-minimal profile so scheduled jobs can call web_search out of the box.

web_search — Multi-Provider Web Search

The web_search tool queries a search engine and returns results. Comis supports 8 search providers — you can configure one or more as a fallback chain.

Parameters

Parameter	Type	Required	Description
query	string	Yes	The search query
count	number	No	Number of results to return (1-10)
deepFetch	number	No	Auto-fetch full content for the top N result pages (0-5, default 0). When > 0, each result includes inline fullContent so you can skip a separate web_fetch call.
country	string	No	2-letter country code for region-specific results (e.g., DE, US). Default: US.
search_lang	string	No	ISO language code for results (e.g., de, en, fr)
freshness	string	No	Filter results by discovery time. Shortcuts: pd (past day), pw (past week), pm (past month), py (past year). Date range YYYY-MM-DDtoYYYY-MM-DD is supported by Brave, Tavily, and Exa. Supported by Brave, DuckDuckGo, Tavily, Exa, and SearXNG; ignored by grok, perplexity, jina.
provider	string	No	Select the first provider for this call (brave, perplexity, grok, duckduckgo, searxng, tavily, exa, jina). Eligible configured fallbacks remain active.

Search Providers

Comis supports 8 search providers. Each has different strengths, and some require API keys while others are free.

Provider	API Key Required	Description
brave	Yes	Brave Search API — fast, privacy-focused
perplexity	Yes	Perplexity AI directly or via OpenRouter — AI-enhanced results
grok	Yes	xAI Responses API — Grok-powered search
duckduckgo	No	DuckDuckGo — no API key needed, uses HTML scraping
searxng	No	SearXNG — self-hosted meta-search engine
tavily	Yes	Tavily — AI-optimized search API
exa	Yes	Exa — neural search for precise results
jina	Yes	Jina Reader Search — content-aware search

Provider Details

Brave Search

Fast, privacy-focused search powered by Brave’s independent search index.

API key: Required. Set SEARCH_API_KEY in your environment.

Best for: General-purpose searching with strong privacy. A good default choice for most setups.

How to get a key: Sign up at brave.com/search/api for a free tier with 2,000 queries per month.

Perplexity

AI-enhanced search results delivered through Perplexity directly or through OpenRouter. Returns synthesized answers alongside traditional search results.

API key: Required. Set PERPLEXITY_API_KEY in your environment. Comis selects the direct Perplexity or OpenRouter endpoint from the key prefix.

Best for: Questions that benefit from AI synthesis, such as “explain the differences between X and Y” or “what are the pros and cons of Z.”

Grok

Powered by xAI’s Grok model via the Responses API. Provides real-time search with AI-generated summaries.

API key: Required. Set XAI_API_KEY in your environment.

Best for: Real-time information and current events. Grok’s training data is frequently updated.

DuckDuckGo

Free search with no API key required. Uses HTML scraping to retrieve results from DuckDuckGo.

API key: Not required.

Best for: A zero-configuration fallback. Works immediately without any setup. Results may be less structured than API-based providers.

DuckDuckGo uses HTML scraping rather than an official API, so result formatting may vary. It is a reliable fallback when API-based providers are unavailable.

SearXNG

A self-hosted meta-search engine that aggregates results from multiple search engines. You run your own SearXNG instance.

API key: Not required, but you need a running SearXNG instance. Set SEARXNG_URL to point to your instance (e.g., http://localhost:8888).

Best for: Privacy-conscious setups where you want full control over the search infrastructure. SearXNG queries multiple engines (Google, Bing, DuckDuckGo, etc.) on your behalf without sending your data to third parties.

Tavily

Purpose-built search API designed for AI agents. Returns clean, structured results optimized for LLM consumption.

API key: Required. Set TAVILY_API_KEY in your environment.

Best for: Agents that need well-structured search results. Tavily is specifically designed for AI-agent use cases and returns content pre-formatted for LLM processing.

Exa

Neural search engine that understands meaning, not just keywords. Finds results based on semantic similarity to your query.

API key: Required. Set EXA_API_KEY in your environment.

Best for: Research-style queries where you need precise, relevant results. Exa excels at finding specific technical content, academic papers, and niche topics.

Jina

Content-aware search and reading service. Retrieves search results with extracted, readable content.

API key: Required. Set JINA_API_KEY in your environment.

Best for: Queries where you need the full content of search results, not just titles and snippets. Jina extracts and returns the readable text from each result page.

Configuring Search Providers

DuckDuckGo is the default primary provider. When it fails, Comis automatically tries the providers for which credentials or an endpoint are configured. Missing-key providers are skipped rather than called with empty authentication. Each eligible provider is attempted sequentially; the first successful result wins.

Programmatic integrations can set fallbackProviders to choose an explicit order. An explicit empty list disables provider fallback. A provider argument on an individual tool call selects the first provider without bypassing the remaining eligible chain.

Environment Variables Reference

Provider	Environment Variable
Brave	SEARCH_API_KEY
Perplexity	PERPLEXITY_API_KEY
Grok	XAI_API_KEY
DuckDuckGo	(none required)
SearXNG	SEARXNG_URL
Tavily	TAVILY_API_KEY
Exa	EXA_API_KEY
Jina	JINA_API_KEY

If every eligible provider fails and the browser tool is enabled, Comis starts the browser and opens a Google Search. Browser fallback never preempts another credentialed search provider.

web_fetch — Page Content Fetching

The web_fetch tool retrieves content from a URL and extracts the readable text using Readability — the same technology behind browser reader modes. This strips away navigation, ads, and other clutter, leaving just the main content.

Parameters

Parameter	Type	Required	Description
url	string	Yes	HTTP or HTTPS URL to fetch
extractMode	string	No	Extraction mode: markdown (default) or text
maxChars	number	No	Maximum characters to return (default: 50000, clamped by configured cap)

How It Works

1. The URL is validated against SSRF (Server-Side Request Forgery) rules
2. The page content is downloaded
3. Readability extracts the main article or content area
4. The clean text is returned to the agent

Important Notes

• SSRF protection — web_fetch rejects URLs that resolve to known localhost, private-network, or cloud-metadata destinations before fetching. Keep the daemon isolated from sensitive internal networks as an additional control.
• Content truncation — Very large pages are truncated to prevent overwhelming the agent’s context window. The tool returns the most important content within the size limit.
• Clean output — The Readability extraction removes navigation menus, sidebars, ads, and other non-content elements. The result is clean, readable text similar to what you see in a browser’s “Reader Mode.”

web_fetch retrieves content from the open internet. Be cautious about fetching URLs provided by untrusted users, as the content could contain misleading or harmful information. SSRF protection prevents access to internal resources, but the fetched content itself is not sanitized for truthfulness.

web_search vs. web_fetch

These two tools complement each other:

Use Case	Tool	Why
Find information about a topic	web_search	Returns multiple search results with titles and snippets
Read a specific web page	web_fetch	Downloads and extracts the full content of a single URL
Research a question	Both	Search first to find relevant pages, then fetch the most promising results for detailed reading

Agents often use both tools together: web_search to find relevant URLs, then web_fetch to read the full content of the best results.

For pages that require JavaScript rendering or interactive navigation (single-page apps, sites behind login screens), use the Browser tool instead — it provides a full headless browser.

End-to-end example: fact-check workflow

A common pattern: a user makes a claim and asks the agent to verify it. The agent searches for primary sources, then fetches the most authoritative result for detailed reading.

You: Verify the claim that the EU AI Act took effect on August 1, 2024.
     Quote the official source.

The agent runs a two-step workflow:

# 1. Search for primary sources
tool: web_search
query: "EU AI Act effective date official Commission"
provider: brave
freshness: py
count: 5

The search returns results from europa.eu, EUR-Lex, and a few news outlets. The agent picks the EUR-Lex link because it is the canonical source.

# 2. Fetch the full content of the official document
tool: web_fetch
url: "https://eur-lex.europa.eu/legal-content/EN/TXT/?uri=OJ:L_202401689"
extractMode: markdown
maxChars: 20000

web_fetch returns the cleaned Readability extract. The agent quotes the relevant article and entry-into-force clause, links the source, and presents the verified answer to the user.

For sources that need fresh data (rate updates, market prices, breaking news) but are not yet indexed by search engines, set deepFetch: 3 on the web_search call — the tool will inline the top three results’ full content, sparing a separate web_fetch.

SSRF protection

Both web_fetch and web_search validate URLs against the SSRF guard before sending any HTTP request. The guard rejects:

• localhost, 127.0.0.0/8, ::1
• RFC 1918 private ranges (10.0.0.0/8, 172.16.0.0/12, 192.168.0.0/16)
• Link-local (169.254.0.0/16) including AWS / GCP / Azure metadata endpoints
• IPv6 unique-local (fc00::/7) and link-local (fe80::/10)

The browser tool applies destination checks to guarded HTTP(S) requests and redirects, and disables page APIs whose traffic cannot pass through that validator. These application controls are not a host firewall, so keep browser automation isolated from sensitive internal networks. See Browser security for the exact boundary.

Related

Browser

Full browser automation for interactive web tasks

Built-in Tools

All built-in tools including web tools

Agent Tools Overview

See all available agent tools

Config Reference

Search provider and web tool configuration options