Transept
Back to changelog
Workflows

The API, and an MCP endpoint for your AI assistant

From the Transept team
Settings → Developer — keys carry their own permissions, rotate in place, and can be revoked one at a time.

Transept has had an API for a while, but only parts of it. This release finishes the job: documents, projects, runs, glossaries, style guides and translation memory are all reachable with a key and no browser session — and the same capabilities sit behind an MCP endpoint, so an AI assistant can drive a localization loop from end to end. Nothing spends words without quoting the cost first.

The second half of the release is quieter and touches every run: a set of house style rules now applies whenever a document has no style guide of its own, and the pass that decides what needs fixing got a sharper rubric.

Added

  • A complete public API. Import documents, run workflows across every language, read the translations back keyed by language, and manage glossaries, style guides and translation memory — all with an API key. Keys are minted in Settings → Developer, carry per-resource read and write permissions, and can be rotated in place: the old secret stops working the moment the new one appears, and the key keeps its name, permissions and history. See Automate Transept with the API and the developer quickstart.
  • An MCP endpoint for AI assistants. Connect Claude Code, Claude Desktop, or Cursor and your assistant gets 41 tools covering the same ground as the API. Connecting is a browser sign-in with nothing to paste; the grant appears as a normal key in Settings → Developer and can be revoked there. Every tool obeys that key’s permissions, and any tool that would spend words answers with an estimate and runs only once you confirm. See Connect your AI assistant (MCP).
  • Team API keys. A team can mint keys of its own, held by the team rather than by whoever created them, so a key outlives the person who made it. Runs started with a team key bill the team pool. Owner-only, and the team needs seats.
  • Webhooks that carry the content. An incoming trigger can now take the text in the request body — raw markdown, HTML, or plain text, or a JSON envelope — and either create a document per delivery or update one fixed document, in which case only the changed part is billed. Retries inside 24 hours are deduplicated. The trigger dialog was rebuilt around it, with a document picker instead of a paste-an-ID box and a copyable example of the call to make. See Trigger a run from another system.
  • An Integrations page. The sidebar’s Integrations item now opens a real page: per-client MCP setup, your connected apps, and incoming triggers next to outgoing webhooks under the names they actually go by. API keys and the reference docs stay together in Settings → Developer.
  • Developer documentation. A quickstart at /developers takes a first key through to translated output, and the help center gained four articles with copyable examples — string catalogs, email templates, markdown, and MCP — under two new sections, Integrations and API & developers. Every article now carries an on-this-page outline.

Changed

  • House style applies when you haven’t written one. A translate, regenerate or fix run on a document with no style guide attached now carries a language-agnostic set of house rules — register, sentence rhythm, punctuation habits, how imagery is handled — so the output reads as written rather than converted. Attaching your own style guide replaces them completely. See Style guides.
  • A sharper editing pass. The step that decides which blocks need work now also flags register drift, sentences that kept the source language’s shape, and departures from your style guide, and tags each finding with a severity and a category. Fixes stay minimal — what the finding names changes, the rest is left verbatim. See Smart Proofread.
  • Starting-point templates translate on Pro. All six templates now use the Pro model for translation and Standard for the editing pass. On the Free plan a template runs on Standard and says so on the card, with the upgrade a click away. See Starting points and templates.
  • One dialog for every run. Translate, Regenerate, Smart Proofread, Proofread with Vision, Translate + Proofread and workflow runs each asked “are you sure?” differently. They now share one shell: the first step says what runs, on what, how long it takes and what it costs. Everything adjustable — model, glossary, style guide, translation memory, guidance, speed — moved behind a Customize button, which a run whose settings were chosen at setup never needs.
  • Runs open on the page you’re looking at. Every run dialog now opens scoped to the current page, with the whole document one click away in the same picker. A first run on the free grant is something you can afford to press.
  • Translation memory is always visible in the run dialog. It used to collapse into “+1 more” on a typical document. It now sits second, right after the model tier, and shows in both states — on, or Off — because a setting that’s off has to be as visible as one that’s on. See Run only what changed.
  • Anything built by AI announces itself. A glossary or style guide produced by a run now raises a dot on the Glossaries and Style guides rows in the sidebar, on the editor rail, and in the mobile Panels hub, plus a catch-up notice for a result you were never told about. The only signal before was a toast in whichever tab happened to watch the run finish.
  • The review walk covers its own run. Opening a result no longer splices in every other result parked days earlier. Finish the last one and the dialog offers the older ones, if any are still waiting.
  • Glossaries and style guides reach a team by moving into it. The separate Share control on those two is retired: a team-owned glossary is already shared with every member, so Share only offered a second route to the same place. Existing shares keep working, and document sharing is untouched. See Glossaries.
  • Export links expire. The file a workflow emails you now stops working after seven days, can be revoked from your side before that, and records how often it was downloaded. The email says so. See Exporting your translation.

Fixed

  • Lists stopped hiding documents. The dashboard fetched a single window of recent documents and split it into personal and team, so a team in active translation could crowd your own documents out of it — 3 cards on an account holding 23. The team page capped at 200 rows against a team of 204, and the document picker only ever searched the 50 most recently touched. All three now page through the full set on the server, folded per language, with the counts coming from the server rather than from what happened to be on screen. Nothing was ever deleted.
  • A clickable toast stays readable. A toast with a button in it no longer fades while you’re reaching for it.
  • Cancelling a run reports what it did. Cancellation over the API and MCP could report an error while actually cancelling; it now answers with the run’s resulting status.

Have a request or hit a snag? Ask Literess in the app, or write to [email protected].