Local-first vs hosted AI summarizers.

"Local-first" and "hosted" sound like clear opposites for AI summarizers, but they're doing more work than the labels suggest. Local-first usually doesn't mean offline — it means the workflow surface (settings, history, keys, exports) stays on your device while the AI generation itself still uses a cloud provider. Hosted means the entire workflow runs on the vendor's infrastructure, including history and (often) source text retention. This page covers what actually differs, what the trust implications are, and when each posture is the right pick.

Four meanings of "local-first"

"Local-first" can mean any of four different things, and products sometimes claim it while only doing one. Local settings: extension preferences (model, output style, custom prompts) stored in Chrome on your device. Local history: summary outputs stored in local storage or IndexedDB on your device, not in a vendor account. Local keys: BYOK provider API keys in Chrome extension storage, not on a vendor server. Local model: the AI itself runs on-device (Gemini Nano, Ollama, llama.cpp). Reduz is local-first in the first three senses; the fourth (on-device model) is a separate category. LocalSum and Chrome's built-in Prompt API summarizer go all four. Most "hosted" AI assistants are none of them.

What "hosted" actually entails

Hosted AI products store settings, history, keys, and prompts on the vendor's infrastructure. The vendor sees what you summarize, when, and how. Monica, Sider, NoteGPT, MaxAI, Eightify, and most all-in-one AI assistants are fully hosted. The trade is convenience: cross-device sync, no key management, faster onboarding. The data path is: your browser → vendor server → AI provider → vendor server → your browser. That extra hop is where the "vendor sees the prompt" question lives. For public content, this is usually fine. For sensitive content, it adds a layer to the trust model.

Why local-first does not mean offline

On-device AI models are a different category from local-first products. Cloud models (GPT-5.5, Claude Sonnet 4.6, Gemini Pro) run on the provider's infrastructure — the request leaves your device by definition. A local-first product can still use a cloud model; the request just goes direct from your browser to the provider instead of through a vendor relay. Offline-only means the model runs on your device — Chrome's built-in Gemini Nano (Prompt API), LocalSum, or local LLMs run via Ollama. Quality is lower than cloud models for most summarization tasks, but the data never leaves your device. Pick offline-only when even the AI provider must not see the source.

The hybrid pattern (and why it works)

A hybrid product offers both a hosted path for quick onboarding and a local-first BYOK path for provider control. Reduz uses this pattern explicitly: Hosted Free (100 monthly credits, no card, prompts through the Reduz relay) for users who want to summarize immediately, plus Your own AI key (your provider key, direct from your browser request, no Reduz relay) for users who care about cost, privacy, or model control. The two modes don't mix — switching between them is one click in settings, and each has its own clear data path. The hybrid pattern beats forcing users to pick one upfront because most users start hosted and graduate to BYOK as their volume grows.

Choosing between local-first and hosted

Choose hosted when: setup friction matters more than provider control, summarization is casual (a few articles a week), you want cross-device history sync, and the source content is public. Choose local-first when: you want BYOK for cost or privacy reasons, you read sensitive documents (drafts, pre-prints, internal reports), you want local summary history that survives account-level decisions by a vendor, or you want narrow Chrome permissions (click-only access vs permission to read every page). For most users, hybrid is the actual right answer — start hosted, switch to BYOK once the workflow stabilizes.

The honest privacy summary

Local-first products reduce the surface area that vendors and intermediate services see; they don't make AI summarization private in an absolute sense. The AI provider still receives the prompt. The provider's data policies still apply (OpenAI and Anthropic exclude API content from training by default; Google AI Studio free-tier policy permits it). For absolute privacy, on-device AI models are the only fully-private answer — at the cost of model quality. For practical privacy with cloud-quality output, local-first BYOK to a provider with API-default no-training policy is the realistic best practice.

Practical checklist

• Decide which of the four "local-first" properties you actually need (settings / history / keys / model).
• For sensitive content, prefer local-first BYOK to a provider with API-default no-training policy.
• For public content and convenience, hosted is usually fine — pick a vendor with reasonable retention policy.
• For absolute privacy, look at on-device models (Gemini Nano, LocalSum, Ollama) — accept the quality trade.
• For most users, hybrid is the right answer: start hosted, move to BYOK when usage grows.
• Read the AI provider's API data policy separately from the vendor app's policy — they're different.
• Verify the local-first claim is real (BYOK request goes direct from your browser, history is in local DB).

Frequently asked questions