Descript MCP with Claude: choose what to edit from competitor moves and AI answers

The Descript connector is a hosted MCP server that works on the Descript Drive you are signed in to. Agent Index lists five tools: import_media, prompt_project_agent, get_job, list_jobs and cancel_job. Descript's help page says Claude can import media, edit with Underlord, manage projects and publish, including filler word removal, Studio Sound, transcription and highlight reels.

The split is simple: Claude writes the instruction and Descript's own agent, Underlord, does the editing. Descript also lists what the directory connector does not cover: generative image or video tools, and Claude Code. For those it points to a custom MCP setup. A third-party review says the connector works on Claude Code too, so check Descript's current help page before you rely on either claim.

Claude can edit what you point it at, but it cannot know what the market is asking for. Compete watches competitors' blogs, changelogs and pages and grades each move with a priority, and Answer Radar records which brands the five engines name and recommend. One plan covers both, at $60 a month. Linkeddit makes no video and edits nothing.

A read-only run on Linkeddit's test project, plausible.io, shows the input. compete_signals_get returned 19 signals for the week of 2026-10-26, including a high priority "Fathom Launches Luna AI Assistant" and five more Fathom launches rated watch. answer_results_get showed Google AI Overviews naming the brand in 15 of 15 usable samples, and ChatGPT in 16 of 18. In the AI Overviews segment youtube.com held 6 of 107 cited-domain votes, so video is a source the engine already reads.

Pick the topic where you are named least, then choose a recording that speaks to it and tell Descript what to cut. On the test project the brand held 10 of 62 vendor votes, a 16 percent share, in the Claude segment, and two segments, Claude on one instrument and Perplexity, returned no usable samples because every sample errored. A good report states that gap.

Claude then sends prompt_project_agent a short instruction built from the report: lead with the claim the engines credit to a competitor, keep the answer in the first thirty seconds, add captions. It imports the recording with import_media first and checks get_job until the job finishes. You approve the brief before it runs, because every edit uses your media minutes and AI credits.

Yes, as an instruction, not as generation. Claude takes a competitor signal, such as a launch with its priority and the reason it matters, and asks Descript's agent to cut a short clip from a recording you already have where a teammate answers that move. Descript's own blog says the connector cannot create AI video, avatars, dubbed audio or cloned voices, so the footage must exist.

The competitor brief was empty on the test project, since compete_brief_get returned brief: null. Report that rather than invent one. Five monitors were set to weekly with a next run of 2026-11-04, so Claude should use compete_signals_get until the brief lands. For a battlecard refresh, the clip is a supplement for sales, and the signal list stays the source of truth.

Cost and quality are the real limits. Descript says imports and Underlord edits use your media minutes and AI credits, and a third-party review says the API and MCP are early access, included with paid Descript plans. A podcaster who tried the workflow wrote on Descript's webinar video: "I'm now out 800 tokens for what should be a simple copy/paste job on barely 12 minutes of content." Descript replied that it should be looked into.

Test with one short recording first, and keep Descript's write tools on approval. Descript's reply to a privacy question says Claude only gets what your own account can already see, permissions split into read and write categories can be set to always allow, needs approval or blocked, and removing the connector cuts access. Claude also cannot watch a folder, so it will not edit a new recording the moment it lands.

Measure the answers, then the views. Record the brand's presence rate and share of voice for the topic from answer_results_get before you publish, then compare the same segments after the next scheduled run. On the test project the Claude segment had 10 usable samples, so one sample moves the rate by ten points.

Re-running answer checks spends credits, and Linkeddit returns an estimate first. Claude should wait for your yes. If the video was published to a page the engines can read, check whether that page now appears among the cited sources. A clip nobody cites is not a failure of the edit, only a sign that the topic needs text and a page as well.