Descript

Drop in footage, a Zoom call or a script and Descript transcribes it, then the transcript becomes the edit surface: delete a sentence and the video loses it. A timeline sits underneath for frame-level work. The AI tools are the reason most people arrive: Studio Sound for noise and echo, filler-word and retake removal, Eye Contact, Green Screen, automatic multicam, captions, clip generation, and translation with dubbing in 30 languages. Transcription covers 25 languages and caption translation 61.

Underlord is the agent layer over all of it. You ask for a rough cut, a highlight reel or a script and it applies the same tools, and Descript's help center still titles it Underlord (Beta).

Every plan carries two allowances per editor. Media minutes count imported or recorded material: 60 minutes on Free, 600 on Hobbyist, 1800 on Creator, 2400 on Business. AI credits count the AI work: a one-time 100 on Free, then 400, 800 and 1500 a month. Nothing on the site says what a given job costs in credits, and the model page notes that premium models use more AI Credits than others.

Top-ups are the escape valve, and they are not available on Free or Hobbyist. Through the API, exhausting either meter returns a 402 error. Free exports carry a watermark and cap at 720p; watermark-free 1080p starts on Hobbyist and 4K on Creator.

Every request goes to descriptapi.com/v1 with a token scoped to one Drive, and import, agent-edit and publish calls return a job_id immediately. You poll it or pass a callback_url that Descript POSTs when the job finishes. There is a CLI, a Zapier integration and an OpenAPI spec. Pricing is the plan you already pay for: API access is available for all paying users at no additional cost.

The MCP server authenticates through your Descript login instead of a token and appears in the Anthropic, ChatGPT and Cursor directories. The limits are published rather than hidden: no download of a rendered file without publishing to a web link, no YouTube URLs as import sources, job history gone after 30 days, and one Drive per connection. Claude's directory connector omits the generative image and video tools.

Descript says it holds SOC 2 Type II through a CPA firm, publishes a trust report, encrypts at rest with AES-256, names a Data Protection Officer and lists every subprocessor it shares data with, including Rev and a self-hosted Whisper for transcription and Hedra for avatar lip sync.

What is absent is as clear. There is no region choice: the terms have you agree that data may be processed in the United States and other countries. There is no on-premise option. SSO, SCIM, audit logs, custom retention and a training opt-out are all Enterprise line items, so a Business team buying 5 seats gets none of them. The only SLA on the published plans covers support response times, and Section 14 of the terms disclaims uninterrupted service.

It fits people whose raw material is talking-head footage: podcast episodes, webinars, sales calls, internal explainers, and the social clips cut out of them. One tool records, transcribes, cleans the audio, cuts, captions and publishes, and the per-person price is low enough that a non-editor can be handed a seat.

It fits badly where volume meets a meter. Forty media hours a month across five Business seats is the ceiling for a team producing daily video, and the credit cost of generated video or avatars is not knowable in advance. Teams that need regional data storage, audit logs or an uptime commitment have to reach Enterprise and a quote. Anyone wanting a finished MP4 out the far end of an automation will still open the app, because the API publishes a web link rather than downloading a render.