What's New

Recent features, improvements, and fixes shipped to Keldura, in plain language. The live page lists every published update newest-first; here are the kinds of changes you'll find:

What's New - Keldura

What's New

Recent features, improvements, and fixes shipped to Keldura — in plain language.

Notes created from MCP show up in the folder

A note added through MCP now appears in its Library folder immediately, and searching for it opens the editor.

Fixed
  • Notes created from a connected MCP client now appear in the Notes folder list even before indexing finishes, and clicking a search hit opens the note.

File Converter accepts images again

PNG, JPEG, BMP, TIFF, and GIF uploads now convert the same way PDFs do.

Fixed
  • File Converter no longer rejects image uploads with a PDF-only error.
  • JPEG, PNG, BMP, TIFF, and GIF files are converted to temporary PDFs during processing, then extracted to Markdown or plain text.

Create Library notes from MCP

Connected MCP clients can now add plain-text or Markdown notes directly to your Library with safe retries.

New
  • The new create_note tool creates an unfiled note for any authenticated API-key owner.
  • Clients can optionally file the note in an owned folder when the account has folder access, and reuse the same idempotency key after a timeout without creating duplicates.

Document cheat sheets stay in the right language

A new cheat sheet from a document now follows your briefing language settings and the video's own language, instead of sometimes picking a third language just because the discussion mentions a country.

Fixed
  • Document cheat sheets use the same language rules as Briefings: languages you chose to keep original, otherwise your default language, otherwise the video's detected language.
  • An English video that talks about Germany no longer comes out as a German cheat sheet when you have not asked for German.

Sharper Grok cheatsheet and cover images

Grok-generated cheatsheets and briefing covers now use the latest Imagine image model, which holds small text more clearly.

Improved
  • When a cheatsheet or briefing cover is generated with Grok, it now uses the newer Imagine Image 2.0 model.
  • OpenAI-generated images and the Briefings A/B compare itself are unchanged — only the Grok arm moved.

Open a Telegram import result directly from Activity

Completed Telegram activity rows that produced a single live document now offer an Open document action in Activity, linking straight to the file. Non-terminal rows, missing files, deleted files, and inactive documents continue to show no action.

Improved
  • A finished Telegram import that produced one document now shows Open document in the Activity log, taking you straight to the transcribed file.
  • The button only appears when the import is complete, failed, or skipped and the resulting file is confirmed live — pending, running, paused, and ready-to-import rows are unaffected.
  • Deleted, inactive, or ambiguous documents never surface the action, so the link is always valid.

Large audio imports stay reliable under slow connections

Importing several big audio or video files at once no longer stalls with a blank HTTP error. Uploads run a few at a time, use smaller chunks, and show a clear retry message when the connection cannot keep up.

Fixed
  • Multi-file audio and video imports now limit how many files upload at once so each chunk finishes before the connection times out.
  • A stalled upload retries once and then fails with a clear message instead of leaving the import stuck as "uploading".

A short video-service outage no longer loses your transcription

When the service we fetch videos from is briefly unavailable, your transcription now waits and tries again instead of failing.

Fixed
  • A transcription whose video service answers with a temporary error is retried automatically, with a growing pause between tries, instead of failing after a few seconds.
  • The activity row shows "Retrying…" while it waits, and the transcription only fails if the service is still unavailable after every retry.

Export transcripts as TXT or SRT

Transcript documents now offer TXT and SRT downloads, preserving stored subtitle sidecars byte-for-byte or generating SRT from timestamped transcripts.

New
  • Transcripts now expose a compact Download menu with TXT and SRT export options.
  • Stored subtitle sidecars are downloaded byte-for-byte without altering timestamps or cue formatting.
  • SRT sidecars are generated on demand whenever every non-empty line carries a valid timestamp.

Transcript documents stay readable

Opening a transcript no longer shows the raw subtitle cue track that was saved beside it.

Fixed
  • When a transcript and its subtitle track finish saving in either order, the document now always opens the readable transcript.

Slow cheatsheets no longer look failed while they finish

When a cheatsheet takes longer than the page's normal wait, Keldura now explains that it may still finish in the background instead of showing a false terminal error.

Fixed
  • Keldura performs one final saved-image check before ending its wait, so a cheatsheet that finishes at the boundary appears immediately.
  • If it is still running, the workspace shows a calm informational notice and keeps Generate available so you can check again shortly.

Transcripts get names and products right more often

We now tell the speech recogniser which names to expect, taken from the video's own title and channel.

Improved
  • People, companies and product names from a video's title and channel are passed to the speech recogniser before it starts, so it spells them the way you would.
  • This applies to video imports transcribed by Grok (title and channel are available there), which is the default on free plans. Other Grok paths that do not yet carry title/channel context are unchanged.

A broken sign-in now gets you back on track

When your saved sign-in is no longer usable, Keldura sends you to sign in again instead of getting stuck.

Fixed
  • If a saved sign-in is revoked or otherwise no longer valid, you are redirected to sign in again after the first affected page load.

X video links copied from the player now work

An X (Twitter) video link ending in /video/1 was turned away as unsupported. It transcribes now.

Fixed
  • Add content and the Import Hub now accept the link X shows once you open a video inside a post, such as x.com/someone/status/1234567890/video/1.
  • Both forms of the link point at the same post, so the saved transcript keeps the post link as its source.

One transcript per import, and one click to read it

Importing a link no longer leaves two near-identical files in your collection, and Activity now takes you straight to what the import produced.

Fixed
  • Importing a video or link used to save two files with the same name — the transcript and the raw subtitle track behind it. Only the transcript is added to your collection now.
Improved
  • A finished import that produced a single document now offers Open document in Activity, which opens the transcript directly instead of leaving you to find it in a collection.
  • The old Open collection button is now Open in library, and it scrolls to and highlights the file the import just added.
  • On a multi-item import, every finished item in View details is now a link to the document it produced.

Folders hold many more documents

Higher per-folder document limits, and your transcripts folder no longer counts against them.

Improved
  • Every plan can keep far more documents in a single folder — the free plan goes from 1 to 25, Starter from 20 to 250, and Pro from 250 to 1000.
  • The automatic Transcripts folder is no longer limited at all, so transcribing another video never runs out of room.
  • An import whose extra file — a caption sidecar, for example — did not fit in the destination folder now finishes with the document you asked for instead of failing.

Keldura Daily runs on our most capable model

Daily curation and story write-ups moved to the strongest model available.

Improved
  • Keldura Daily picks its lead stories and writes their summaries using our most capable model.
  • Everything else keeps running on the fast model, so nothing else got slower.

The highest reasoning level is selectable again

Bots on the newest models were missing their top reasoning level in the picker. It is back.

Fixed
  • The reasoning level picker for bots, Telegram bots and admin model assignments now offers every level the selected model actually accepts.
  • Models released after the picker was written were silently capped at High even when they supported more.

A better model behind every AI feature

Summaries, titles, translations, tags and briefings now run on a newer, more capable model.

Improved
  • Summaries, cheatsheets, suggested questions, transcript titles, translations, monitor tags and briefing editions all moved to a newer model.
  • Free accounts get the newer model too — it replaces the older one they were limited to.
  • Nothing changed about how you use these features; the same actions just produce better results.

Chat costs fewer credits

Our AI provider cut its prices, and we passed the cut straight through to your credit balance.

Improved
  • A chat turn on the default model now costs 1 credit instead of 3, so your monthly allowance goes about three times further.
  • Chats on the higher-capability Terra model dropped from 7 credits to 6 per turn.
  • Every other model keeps the price it had — only the models whose provider pricing actually changed were repriced.