Adding Research Capabilities

Now let's give your research agent some superpowers by adding specialized tools for web research.

Why Add These Tools?

The built-in WebSearch and WebFetch tools are great for basic tasks, but they have limitations. WebSearch returns snippets and may not always give you full content. WebFetch can struggle with JavaScript-heavy sites or complex page structures.

This is where specialized services come in:

Firecrawl is a web scraping service that turns any website into clean, LLM-ready markdown. It handles JavaScript rendering, removes navigation and ads, and extracts just the content you need.

Tavily is an AI-powered search API built specifically for LLMs. It understands context better than traditional search and returns highly relevant sources with clean summaries.

Together, they create a powerful workflow: Tavily finds the best sources, and Firecrawl extracts clean content from them.

Prerequisites

You'll need API keys from both services:

• Firecrawl API key from firecrawl.dev (500 credits/month free)
• Tavily API key from tavily.com (1,000 credits/month free)

You should have also completed the previous sections of this tutorial.

Step 1: Add Skills to vm0.yaml

Open your vm0.yaml and add the skills section:

version: "1.0"

agents:
  my-researcher:
    framework: claude-code
    instructions: AGENTS.md
    skills:
      - https://github.com/vm0-ai/vm0-skills/tree/main/firecrawl
      - https://github.com/vm0-ai/vm0-skills/tree/main/tavily

The skills section tells VM0 to load these integrations when composing your agent.

Step 2: Update Agent Instructions

Now update your AGENTS.md to use these new tools:

# Deep Research Agent

You help me research topics by gathering information from multiple sources and creating a summary report.

## What to do

When I give you a research topic:

1. Create a folder based on the research topic to keep things organized
2. Using tavily figure out 3-5 good sources to look at
3. Using firecrawl visit each source and grab the important info
4. For each source, write a research report about what you found from that source
5. Combine everything into one report called `report.md` that summarizes all the findings

That's it! Keep the files simple and readable.

The key changes are explicitly telling the agent to use tavily for search and firecrawl for content extraction.

Step 3: Compose Your Agent

Recompose your agent to include the new skills:

vm0 compose vm0.yaml

You'll see output showing the skills being downloaded and uploaded:

Uploading instructions: AGENTS.md
✓ Instructions uploaded: 15f0ffee
Uploading 2 skill(s)...
  Downloading: https://github.com/vm0-ai/vm0-skills/tree/main/firecrawl
  ✓ Skill (unchanged): firecrawl (119bf88e)
  Downloading: https://github.com/vm0-ai/vm0-skills/tree/main/tavily
  ✓ Skill uploaded: tavily (f1cc346f)

Skills require the following environment variables:

  Secrets:
    FIRECRAWL_API_KEY        (new) <- firecrawl
    TAVILY_API_KEY           (new) <- tavily

? Approve 2 new secret(s)? › (Y/n)

VM0 is asking you to approve the new secrets required by these skills. This is a security confirmation since the secrets are implicitly passed through the skills. Type y and press Enter to continue.

Step 4: Store Your API Keys

Before running, store your API keys on the platform (one-time setup):

vm0 secret set FIRECRAWL_API_KEY fc-xxx
vm0 secret set TAVILY_API_KEY tvly-xxx

This stores your secrets securely on VM0's servers. They'll be automatically loaded whenever you run your agent.

Step 5: Run Your Agent

Now let's test it out:

vm0 run my-researcher --artifact-name artifact "research the latest developments in WebAssembly for 2025"

Since your secrets are stored on the platform, you don't need to pass them every time!

This time it should work! You'll see the agent:

1. Use Tavily to find relevant sources
2. Use Firecrawl to extract content from each source
3. Create individual research reports
4. Combine everything into a final report

The research quality will be noticeably better compared to using basic web tools.

What Just Happened?

The skills system integrated Firecrawl and Tavily into your agent's capabilities. When you composed the agent, VM0:

1. Downloaded the skill definitions from GitHub
2. Built the $HOME/.claude/skills directory structure to mount into the sandbox
3. Injected the API keys from your environment into the sandbox securely

During execution, Claude Code could see these tools in its available toolkit and used them according to your instructions. The combination of AI-powered search and clean content extraction produced much higher quality research.

To learn more about how skills are mounted into the sandbox, see the Volume documentation. For details on how environment variables and secrets work, check the Environment Variables documentation.

Next Steps

Your research agent is now much more capable! In the next part, we'll set up automatic scheduling so your agent can run research tasks on a recurring basis.