Building a Research Assistant with Claude Code and Web Search

Claude Code is primarily a coding agent, but it spends a lot of time doing research — looking up how things work, evaluating options, checking current best practices. Without web access, that research is limited to its training data. With AgentPatch, it can actually search.

Why This Matters

Technical research during development usually looks like: tabbing out to Google, reading a few pages, coming back, and trying to summarize what you found into a prompt. With web search available as a tool, Claude Code does that loop itself.

This is useful beyond just “what version is this package?” — it handles more open-ended research tasks: comparing architectural approaches, checking whether a library is still maintained, understanding the security implications of a dependency. The agent can search, read, and reason about the results without you leaving the terminal.

Setup

Install the AgentPatch CLI (zero dependencies, Python 3.10+):

pip install agentpatch

Set your API key:

export AGENTPATCH_API_KEY=your_api_key

Then use it:

ap search "web search"
ap run agentpatch google-search --input '{"query": "test"}'

Get your API key from the AgentPatch dashboard.

Run this command to add AgentPatch as an MCP server:

claude mcp add -s user --transport http agentpatch https://agentpatch.ai/mcp \
  --header "Authorization: Bearer YOUR_API_KEY"

Replace YOUR_API_KEY with your actual key from the AgentPatch dashboard. Claude Code discovers all AgentPatch tools automatically.

Example

You’re evaluating options for background job processing in a Node.js app:

“I need to add background job processing to this app. What are the main options right now and which would you recommend for a small Postgres-backed service?”

Claude Code searches before answering:

Searching Google for "Node.js background job queue Postgres 2026"...
Searching Google for "BullMQ vs pg-boss vs graphile-worker comparison"...

For a Postgres-backed service, the practical choices are:

- **pg-boss**: Jobs stored in Postgres, so no additional infrastructure. Simple API,
  good for lower volumes. Active development.
- **graphile-worker**: Also Postgres-native, more feature-rich, used in production at
  scale. Slightly more setup.
- **BullMQ**: Backed by Redis, so you'd need to add Redis. Better throughput if you
  need it, but that's extra ops overhead.

Given you're already on Postgres and this sounds like a smaller service, I'd start with
pg-boss. Want me to install it and set up a basic job queue?

Claude Code ran a targeted search, compared current options, and made a concrete recommendation. That’s the research loop compressed into a single step.

Wrapping Up

Adding web search to Claude Code makes it a more useful research partner during development, not just a code generator. AgentPatch sets it up in one config block, and the connection gives access to the full marketplace — image search, email, maps, and more. See agentpatch.ai.

← Back to blog