Introducing Parallel Search Turbo: the fastest and lowest-cost search for AI. Learn more.[Learn more.]

August 18, 2026

# How to add web search to a local AI agent (Ollama, LM Studio, Pi, and Goose)

A local model has no idea what happened after its training cutoff, and most search APIs would put a bill back into your zero-cost stack. Here's how to add free web search to agents running on Ollama and LM Studio, via LM Studio, Goose, Pi, and Cline, plus the honest privacy caveat.

Tags:Guides
Reading time: 4 min

Running a model locally solves cost, privacy, and control in one move, and creates a new problem in the same move: the agent's knowledge ends at its training cutoff, and a local machine has no web index. Ask a local Gemma or Llama about a library released last month and you get confident fiction. Web search is the missing organ, and the trick is adding it without reintroducing the API bill your local stack exists to avoid.

The good news: the Parallel Search MCP[Parallel Search MCP] is free with no API key or account, so a fully local agent gains real web search and page fetching for exactly $0. Disclosure: it's ours, which is why this guide uses it; the setup patterns below work with any hosted MCP server if you'd rather bring a different one.

## One caveat before setup

Adding web search punctures pure locality: your search queries and fetched URLs go to the search provider, even though your prompts, documents, and the model's reasoning stay on your machine. For most people that's the right trade, since a query is far less sensitive than a conversation. If your threat model forbids even queries leaving the machine, stop here; no hosted search fits, and a local crawler won't give you a web-scale index.

## LM Studio

LM Studio runs models locally and speaks MCP natively, with one-click install links (the lmstudio:// deeplink on our docs page) or a manual entry in its mcp.json. Once added, any tool-capable local model can call web_search and web_fetch like any other tool.

{
  "mcpServers": {
    "Parallel Search MCP": {
      "url": "https://search.parallel.ai/mcp"
    }
  }
}

## Goose

Goose, Block's open-source agent, runs against local models and adds MCP servers as extensions. Use the one-click goose:// install link from our docs, or add an extension manually with type streamable_http and the URL https://search.parallel.ai/mcp. It works without a key.

## Pi (with Ollama)

The Pi coding harness pointed at an Ollama model is the purest zero-cost stack, and we've published a complete walkthrough: building a fully free CLI agent with Pi, Ollama, Gemma 4, and Parallel[building a fully free CLI agent with Pi, Ollama, Gemma 4, and Parallel]. Pi has no built-in MCP support, so either install pi-mcp-adapter and add the server to .mcp.json (details in our Pi MCP guide[Pi MCP guide]), or skip MCP and install the Parallel CLI as a Pi skill, which suits Pi's shell-first design.

## Cline (with Ollama)

Cline in VS Code can drive Ollama-served models, and it treats remote MCP servers as stdio processes via the mcp-remote wrapper:

{
  "mcpServers": {
    "Parallel Search MCP": {
      "command": "npx",
      "args": ["-y", "mcp-remote", "https://search.parallel.ai/mcp"]
    }
  }
}

## Two tips for small local models

First, keep the tool count low. Small models degrade at tool selection much faster than frontier ones, so a local agent with two well-described tools (web_search, web_fetch) will outperform the same model juggling twenty. Second, lean on search excerpts. The Parallel server returns excerpts dense enough to answer from directly, which matters doubly when your model's context window is 8k or 32k rather than a million; teach your agent, via system prompt or skill, to fetch full pages only when the excerpts fall short.

## The zero-dollar stack, summarized

LayerFree optionCost
ModelOllama or LM Studio (Gemma, Llama, Qwen)$0
HarnessPi, Goose, Cline, or LM Studio's chat$0
Web search + fetchingParallel Search MCP, keyless$0
Optional: deeper researchParallel Task MCP with free account$5/month free credit

## Frequently asked questions

**Does Ollama support MCP directly?** Ollama serves models; the MCP connection belongs to the harness in front of it (Pi, Goose, Cline, LM Studio, and others). Point any of them at your Ollama endpoint and add the server there.

**Will a small local model use search tools well?** Recent small models handle tool calling fine when the tool count is low and descriptions are clear. When they stumble, the cause is usually too many tools rather than tool calling itself.

**Is the free tier really enough?** For personal use, yes; that's what the anonymous rate limits are sized for. If you outgrow them, a free account adds $5 in recurring monthly credits before anything costs money.

## Give your local model the internet

Pick your harness above, paste one config block, and ask your local model something that happened last week. The whole experiment is free, and the Pi + Ollama walkthrough[Pi + Ollama walkthrough] shows the complete build if you're starting from nothing.

Parallel avatar

By Parallel

August 18, 2026

## Related Articles8