Connect your AI assistant (MCP)

MCP (Model Context Protocol) is the standard AI assistants use to call outside tools. Transept publishes one endpoint for it, https://app.transept.ai/api/public/v1/mcp, and an assistant connected to it gets the same capabilities a script gets from the API, packaged as tools it can call on its own: create a document, price a run, start it, poll it, read the translations back.

The practical difference from copy-pasting text into a chatbot: the assistant works on your real documents, with your glossaries, your style guides, and your translation memory applied. The results land in Transept, where you can review and export them like any other work.

Connecting uses standard OAuth: add the URL to your client, and the first connection opens a browser window asking you to approve the access. Approve it and the client is linked; there is no key to paste. Ready-made snippets for each client are on the Integrations page in the app.

Not for interactive clients: OAuth covers them. Keys are for machines that can’t open a browser. Either way the connection ends up under Settings → Developer as a revocable entry, so disconnecting is one click no matter how it was made.

The tools cover the loop end to end, so a single conversation can take a document from import to reviewed translations:

• Import content: paste-in documents, or a keyed string catalog with contexts and placeholders.
• Price any run for free, then run a workflow across every target language at once, and poll progress per language.
• Create and extend glossaries, read style guides, and query translation memory before translating.
• Read results back: per block for prose, or keyed by string id for catalogs.
• Register webhooks so your own systems get pushed a signed event when a run finishes.
• What stays outside, by design: file uploads and binary exports (DOCX, XLIFF, TMX) live on the REST API, and key management stays a human action in the app; an assistant can never mint itself a key.

Three guardrails, all enforced on Transept’s side rather than trusted to the assistant:

• Permissions. Every tool call is checked against what you granted: the scopes of the key, or of the OAuth consent. A call outside them is refused with the missing permission named.
• Estimate, then confirm. Anything that would spend words first returns the word estimate and stops. The run only starts when the assistant confirms explicitly, which means you see the price in the conversation before it happens. Review gates in your workflows still pause the run; continuing one takes an explicit approval, from you or from the assistant at your instruction.
• Revocation. The connection can be cut from your side at any moment: a revoked assistant’s next call is refused, mid-conversation or not.

Yes: pair the connection with a webhook trigger and the loop closes. Your email platform, CMS, or a script POSTs content to the trigger URL (the request body becomes the document), the workflow translates it, and a webhook or the assistant reads the finished translations back: ten languages out for one template in, with nobody opening the editor.

This is the pattern behind email template localization and continuous string catalog work: the platform pushes what changed, Transept re-translates only that, and your pipeline pulls the results. Your part is the review — everything mechanical is delegated.