You ask your AI assistant "what's the latest on the school board vote?" and get a confident-sounding answer — from six months ago. You ask "which airline has the cheapest flight to Rome next week?" and the response is a generic paragraph about how to search for flights. The model doesn't know because it can't see the web.
Meggy's web search changes that. Flip the toggle, and every question can pull fresh results from the live internet — with sources, citations, and a direct link back to each page. For bigger questions, Deep Research goes further: it breaks your question into sub-queries, searches them in parallel, cross-references the findings, and writes you a sourced report. All of it happens locally, and you control every step.
When web search is enabled for a chat, Meggy calls out to real search engines behind the scenes. You get actual search results — not hallucinated URLs, but real links with real snippets from real pages.
Meggy supports six search providers. You can configure one or several — the system tries each in order until it gets results:
| Provider | Free Tier | What Makes It Stand Out |
|---|---|---|
| Tavily | 1,000/mo | Rich snippets, recommended default |
| Brave Search | 2,000/mo | Privacy-focused, no tracking |
| Serper | 2,500/mo | Google SERP results via API |
| SearXNG | Unlimited | Self-hosted meta-search engine |
| Jina AI | 1,000/mo | Reader-style content extraction |
| Perplexity Sonar | — | LLM-powered search with citations |
To keep things fast and avoid burning through your API quota, Meggy caches search results for 15 minutes. If you (or the AI) ask the same question twice within that window, the cached results come back instantly.
Sometimes a quick search isn't enough. "Compare the pros and cons of heat pumps vs. mini-splits for a 1,200 sq ft home in the Northeast" — that's not a single-search question. That's a research project.
Deep Research handles it automatically with a 6-phase pipeline:
An LLM breaks your question into 3–7 focused sub-questions. For the heat pump example, it might generate: "What are the upfront costs?", "How do they perform in sub-freezing temperatures?", "What's the maintenance schedule?"
Each sub-question is searched in parallel across your configured providers. This phase uses the same fallback chain as regular web search.
For the top-ranked results, Meggy fetches the full page content — not just the snippet — so the AI has complete context to work with.
The AI compares findings across all sources so far. Where do sources agree? Where do they conflict? This step catches contradictions and highlights consensus.
Everything comes together into a comprehensive report with inline [1] [2] [3] citations linking back to each source. You can click any citation to open the original page.
If you asked for it, the entire research package — every source URL plus the final report — is saved to a Vault folder so you can reference it later.
| Mode | Sub-Questions | Best For |
|---|---|---|
| Quick | 3 | Fast overviews, simple comparisons |
| Thorough | 5–7 | In-depth analysis, multi-faceted topics |
Both regular web search and deep research can save results directly to your Vault — Meggy's local document store:
Web Search/school board vote Mar 2026 and batch-ingests each result URL as a Vault documentResearch/heat pumps vs mini-splits Mar 2026 folder, ingests all source URLs, and stores the synthesis report as a searchable noteOnce saved, you can ask follow-up questions weeks later — "what did that research say about maintenance costs?" — and the Vault search finds it instantly.
A globe icon (🌐) in the message input lets you enable or disable web search per conversation. Each chat remembers its setting, so your casual chats stay fast while your research chats stay informed.
When the AI searches the web, you see an expandable card showing each result with its favicon, domain name, and a snippet preview. Click any result to open the original page.
Answers include numbered badges like [1] [2] [3] that link directly to source URLs. A citation bar below the message collects all sources in one row for easy scanning.
During deep research, a live progress card shows which phase the engine is on, how many sub-questions have been searched, and what it's currently doing — so you're never left wondering.