New: monitor which AI crawlers actually visit your site with Sona Agent Analytics  |  See the platform →

MistralAI-User

Mistral's user-triggered fetcher. One of three separate Mistral tokens, cleanly split by purpose.

Reference reviewed by Sona

OperatorMistral AI
PowersLive page fetches for Mistral's assistant
PurposeUser-triggered fetch
User-agent tokenMistralAI-User
Respects robots.txtYes

MistralAI-User fetches a page when a user of Mistral's assistant asks a question that requires visiting it, and Mistral notes the assistant will "include a link to the source in its response." Mistral states this token is not used for automatic crawling or for generative AI training - it exists only to serve direct user actions.

Mistral's crawler design is the cleanest separation of concerns of any operator in this directory, and it now runs three tokens rather than one: MistralAI-User for user-triggered fetches, MistralAI-Index for building the search index, and MistralAI-Training for building training datasets. Each has its own purpose statement, its own robots.txt token, and its own published IP feed.

That granularity makes a precise policy possible in a way most vendors do not allow. You can disallow MistralAI-Training to stay out of the training corpus while allowing MistralAI-Index and MistralAI-User to remain findable and citable - the same shape as the GPTBot-versus-OAI-SearchBot split, but with the index and live-fetch roles also separated.

Anyone whose robots.txt names only MistralAI-User is covering one third of Mistral's crawling. The Index and Training tokens fall through to whatever the wildcard group says, which for a training opt-out is usually the wrong answer.

How MistralAI-User behaves

  • Single-page fetches tied to one user's question; Mistral states it does not crawl automatically.
  • Documented as never used for generative AI training - that role belongs to the separate MistralAI-Training token.
  • Cites its source in the assistant's response, so an allowed fetch can produce a visible link.
  • One of three Mistral tokens, each with its own published IP feed.

Full user-agent string

Mozilla/5.0 AppleWebKit/537.36 (KHTML, like Gecko; compatible; MistralAI-User/1.0; +https://docs.mistral.ai/robots)

Allow MistralAI-User

Users asking Mistral's assistant about your page get an answer grounded in the live page with a source link - and Mistral states this token never feeds training.

User-agent: MistralAI-User
Allow: /

Block MistralAI-User

You want no programmatic access from Mistral's assistant, including user-initiated fetches.

User-agent: MistralAI-User
Disallow: /

How to verify MistralAI-User

Published IP range feed

Mistral publishes a per-token IP feed - mistralai-user-ips.json for this bot, with separate files for the Index and Training crawlers. That one-file-per-purpose structure is unusually disciplined and makes auditing straightforward: if you have disallowed MistralAI-Training, traffic verified against the training feed is a compliance problem you can actually evidence, rather than an ambiguous user-agent string.

Check an IP against this bot

Commonly confused with MistralAI-User

ChatGPT-User

Same role at OpenAI, but a different guarantee: Mistral documents robots.txt compliance for this token, while OpenAI says robots.txt may not apply to ChatGPT-User.

cohere-ai

Another European-vendor token that is often miscategorised as a training crawler when it actually serves user-initiated prompts.

MistralAI-User FAQs

How many crawlers does Mistral run?

Three, each separately documented: MistralAI-User for user-triggered fetches, MistralAI-Index for search indexing, and MistralAI-Training for training datasets. Every one has its own robots.txt token and its own published IP feed.

Does MistralAI-User feed model training?

No. Mistral states this token is not used for automatic crawling or generative AI training. Training collection is handled by the separate MistralAI-Training token, which is the one to disallow if that is your concern.

My robots.txt only names MistralAI-User. Is that enough?

No - it covers one of three tokens. MistralAI-Index and MistralAI-Training will fall through to your wildcard group, which for a training opt-out is usually not what you intended.

Can I be citable in Mistral without being training data?

Yes, and the separation is cleaner than at most vendors. Allow MistralAI-Index and MistralAI-User, disallow MistralAI-Training.

Can MistralAI-User read your page right now?

Test any URL and see exactly what AI crawlers receive.

Check my site