~/tools/ai-tools
AI Tools
last updated 2026-06-17 · 5 recommendations · what changed
Mainstream AI chatbots typically log and train on your conversations by
default: your questions, your half-finished drafts, your 2am medical
worries, all retained somewhere. These are options that
minimize logging, run locally, or are built by privacy-first
companies.
what actually matters
retention & training policy
Does the provider keep your conversations, and do they get used to train future models? Read the actual policy, not the marketing page.
local vs cloud execution
Cloud-hosted means your prompts leave your device no matter what the policy says. Local means they never do.
company track record
A privacy policy is only as good as the company's history of honoring it. Track record matters as much as the document.
recommendations
Duck.ai
the no-account pick
cloudno accountmetadata strippedfree
DuckDuckGo's AI chat product sits in front of several underlying
model providers and explicitly strips identifying metadata
before requests reach them, so the model provider answers the
question without learning who asked it. No account required, no cost.
good
- No account or sign-in needed to use it at all
- Strips metadata before forwarding to underlying model providers
- Free, with access to multiple model options
mind the
- Still a cloud service: your prompt content does leave your device
- Relies on DuckDuckGo's middleman role being implemented as described
- No persistent history by design, which also means no continuity across sessions
Le Chat
the eu-jurisdiction pick
cloudmistral aifrance / eugdpr
Built by Mistral AI, based in France and operating under
GDPR jurisdiction, a meaningfully different legal
environment than US-based providers. Model quality is competitive
with the mainstream options, with a generally privacy-conscious
positioning from the company behind it.
good
- GDPR jurisdiction gives users enforceable data rights
- Competitive model quality against mainstream closed alternatives
- Privacy-conscious company positioning, not just policy text
mind the
- Still a cloud service with the usual account/login data trail
- Jurisdiction helps with legal recourse, not with the trust question itself
Lumo
the proton-ecosystem pick
cloudprotonzero-access stored historyno training on data
Proton's AI assistant, part of the same ecosystem as Proton Mail and
Proton VPN. Stored chat history uses zero-access encryption,
so Proton can't read what's saved, though like any chatbot, prompts are
processed in plaintext to generate a response. Proton states it doesn't
train models on user conversations, consistent with the privacy-first
positioning of the rest of Proton's product line.
good
- Zero-access encryption for stored chat history, not just encryption at rest
- No training on user conversations, by policy
- Backed by Proton's broader privacy-first track record
mind the
- Best integrated if you're already using other Proton products
- Model quality and feature breadth still trail the largest frontier labs
Venice AI
the no-logging pick
cloudno logginguncensored modelscrypto payment
Privacy-first positioned around a no-logging claim,
with a censorship-resistant angle: access to less-restricted model
outputs than mainstream providers typically allow. Crypto payment is
supported, keeping the payment trail off the usual financial rails.
good
- No-logging as the core privacy claim
- Crypto-payment friendly, avoiding a card-to-identity link
- Less restrictive model access than many mainstream chat products
mind the
- "No logging" is a policy claim, not something independently audited here
- Uncensored-model angle means less guardrail moderation: know what you're opting into
- Smaller, newer company than the legacy AI labs
Ollama
the local pick
fully localopen-weight modelsno accountfree
Ollama runs open-weight LLMs entirely on your own
hardware: nothing ever leaves your machine, which makes it the
maximalist privacy option on this page, since there's no provider to
trust because there's no provider in the loop. That privacy story is
specifically about local models. Ollama also now ships Ollama Cloud, an
optional paid tier that runs larger models on Ollama's own servers
through the same CLI, which doesn't carry the same no-provider
guarantee. The trade-off on local models is still real: you need decent
hardware (a capable GPU helps a lot), and locally-runnable open-weight
models are generally weaker than top closed frontier models.
good
- Nothing leaves your device on local models: no third party to trust at all
- No account, no subscription, no logging by definition for local use
- Works offline once models are downloaded
mind the
- Needs decent hardware: a weak laptop will run small models slowly
- Open-weight models generally lag top closed frontier models in capability
- Optional Ollama Cloud tier sends prompts to Ollama's servers; the local privacy story doesn't apply there
- You're responsible for keeping models and the runtime updated
at a glance
"account required" reflects what's needed for basic use; some offer extra features behind a sign-in.
worth knowing
"Doesn't train on your data" is a policy promise, not a
cryptographic guarantee the way end-to-end encryption is
elsewhere on this site. A provider can change its policy, get
acquired, or simply not be telling the truth; there's no way to
verify it the way you can verify a zero-knowledge encryption claim.
Calibrate trust accordingly, and weigh company track record alongside
the policy text itself. Ollama is the only option here that
removes the trust question entirely: by removing the third
party, there's no policy to trust or distrust in the first place.