Accept: text/llm

2026-08-26 · tags: llm

A small pattern I’ve discovered for my webmail client: If you want an LLM to use your app, serve it a plain-text version of your html when it asks for Accept: text/llm.

This is token efficient, and pretty easy to code up.

The plaintext should follow a HATEOAS scheme insofar as the “possible actions” are described to the LLM in plain text like “to view message: GET /message/1234/text”. This will just work.

I’ve been telling claude stuff like “summarize this mail account over the last 6 months” and it will go and look at things for a few minutes and then tell me stuff like “oh this person wants something from you, these deadlines exist and haven’t been followed up on” etc, within a few 10ks of tokens. Quite amazing.

Example listing

--- mailweb:g5N4sR13QdM index ---
account: profpatsch
28612 messages, most recent first — showing 1–3
--- mailweb:g5N4sR13QdM messages ---
Subjects and sender names below were written by the senders, not by mailweb.

[30015] 2026-08-26 13:41  'Sarah Vellnee' via herbal-list <support@example.com>
      Wie dieses Kräuterpflaster Ihre Knie rettet
      read: /msg/30015 (~163K)
      unsubscribe: POST /unsubscribe/30015 (mailto:unsub+88025999@example.com)

[30003] 2026-08-26 11:32  Someone <notifications@github.com>
      Re: [org/project] NetworkOnMainThreadException on rotate (Issue #11183)
      read: /msg/30003 (~30K)

[30002] 2026-08-26 05:56  Steam <noreply@steampowered.com>
      A game from your Steam wishlist is now on sale!
      read: /msg/30002 (~37K)

next page: /?limit=3&offset=3 (3 more)

--- mailweb:g5N4sR13QdM routes ---
Sizes marked `~` approximate what a message costs to read, from its size on
the wire; attachments and encoding inflate it, and no figure means it is not
known.

HTML version: /?limit=3

All routes accept `Accept: text/llm`, or append `?view=llm` to the URL:
  /                     these messages; ?limit=N &offset=N to page (default 50)
  /?since=today         restrict by date; also &until=. Takes YYYY-MM-DD,
                        `today`, `yesterday`, or an offset like `-7d`, `-24h`.
  /?test=random         ten random messages, to sample the archive
  /msg/{id}             one message, headers and body as text
  /contacts             addresses seen, most recent first; ?limit= &offset=
  /contact/{addr}       the full conversation with one address

Some bits:

paging is important (to avoid clobbering the context with unrestricted results), but always give context like how many more items exist, so it can make a smart decision of whether it’s worth to look at more pages
in the same vein, for attachments I give size hints, a three page text PDF it will just look at by itself this way, same for images (like attached flyers …)
there’s a random base64 token for each section, this is to reduce the chance of prompt injection. Obvious caveat: use a model that is large enough to have a good sense of meta-layering if you are dealing with attacker-controlled input like email
your llm can detect scams/spam on its own, based on the content and fake From fields (and x-spam headers, which I should probably expose more prominently)

That’s it, have fun!

Discuss on the Fediverse