> ## Content Index
> Fetch the complete content index at: https://www.implicator.ai/llms.txt
> Use this file to discover other available public pages before exploring further.

# Which DAW Is Best for Bringing AI Agents Into Your Workflow?
- URL: https://www.implicator.ai/best-daw-ai-agents-reaper-pro-tools/
- Published: 2026-11-08T21:40:28.000Z
- Updated: 2026-11-08T21:40:28.000Z
- Description: REAPER, Pro Tools and Resolve now take orders from Claude, but the safety rails differ by editor. Inside: a REAPER test of open-source TypeWhisper scripts that cut sound bites from a transcript, six editors compared on agent control and price, and why Adobe offers no agent route into Audition.
- Author: Marcus Schuler
- Tags: Implicator PRO Briefing, analysis

*Implicator PRO Briefing / 9 Nov 2026*

On Oct. 31, 2026, we loaded a raw interview into REAPER on a Mac: 1 hour 52 minutes 45 seconds of conversation, a 1.95 GB WAV copied straight off the recorder. We selected the clip and ran a small Lua script that handed the audio to TypeWhisper, a transcription app on the same computer. REAPER stayed usable while the job ran. When it finished, 1,067 labeled regions lay across the waveform, each a stretch of the interview with its words attached, none overlapping.

The script had not existed that morning. Marco Hillger, who develops TypeWhisper, sent the first version back 59 minutes after we suggested the idea by email. By the next day there were three scripts, published on GitHub under an open-source license, and the third did the work radio producers spend their afternoons on. It read a manuscript, found each quoted passage in the recording and cut it as a finished sound bite, placing every cut in the pause between words. We wrote that third script, and the project credits us as a contributor, so on REAPER we are not a neutral tester.

Radio producers and podcasters choosing an editor now face the same question across the market. Between April and October 2026, the editors they use began accepting instructions from AI agents such as Claude, and each draws a different line around what an agent may touch in an interview's source audio.

The routes into those applications differ. Ardour and Blackmagic, the company behind DaVinci Resolve, ship connections of their own. Descript and Riverside offer theirs through a cloud account. REAPER and Pro Tools depend on outside developers, and one of the Pro Tools add-ons turns the program's windows into text and sends that description to a chatbot. Adobe has opened Photoshop and Premiere to an outside chatbot and left Audition off its list. Hillger's work is another kind again: a transcription app wired into the timeline, with the engine, on-device or in the cloud, chosen by the person at the desk.

The research for this piece turned up no head-to-head test of these editors on the same interview. The TypeWhisper tests ran on a single machine, and they tested transcription laid onto the timeline and scripted cuts, the groundwork an agent needs, not an agent cutting an interview. It measured no review time saved. What follows rests on that trial, on vendor documentation and community code, and on a long forum argument among the people who make Ardour.

The short version

- REAPER's agent route is community code: TwelveTake's 180-tool server, written for music production and tested through REAPER 7.79.
- TypeWhisper's REAPER scripts, open source since Nov. 1, turn an interview into timeline regions and cut sound bites from a marked transcript or manuscript; they are scripts, not an agent.
- Pro Tools has native speech-to-text but no agent from Avid; Adobe offers no agent route into Audition at all.
- Descript, Resolve Studio and Ardour ship their own connections, limited by publish-first downloads, desktop-only Studio and opt-in local models.

AI-generated summary, reviewed by an editor. [More on our AI guidelines](https://www.implicator.ai/about/).

## Six months of agent connections

Most of the new links run on the Model Context Protocol, or MCP. It is a standard plug that lets an AI assistant call an application's functions directly: list what a project contains, drop a marker, start a render. An assistant can reach only what the application has chosen to wire behind that plug.

On April 28, Adobe launched a connector for its creative applications, and Avid shipped Pro Tools 2026.4 with new transcription features.

On May 19, 2026, Ardour 9.5 shipped what its [release notes](https://ardour.org/news/9.5.html?ref=implicator.ai) call "a community-contributed MCP server for controlling Ardour from a locally running large language model." The open-source workstation made it opt-in, behind a checkbox marked experimental, and named accessibility for people with visual impairments as a practical use. The setup guide it points to is written for a model running on the same computer.

On Sept. 8, at the IBC broadcast trade show, Blackmagic announced [DaVinci Resolve 21.1](https://www.blackmagicdesign.com/media/release/20260908-03?ref=implicator.ai), which "now supports integration with AI assistants such as Claude, Claude Code and ChatGPT Codex." Resolve 21.1 is a free download. The assistant feature is listed only for the paid edition on desktop computers.

On Oct. 27, Forte, which already sold session-preparation software for Pro Tools, released [Braus](https://www.forte-ai.com/blog/braus-for-pro-tools?ref=implicator.ai) in beta. It is an agent that connects a customer's own chatbot subscription to Pro Tools and works in the background while the user carries on.

By the end of October, two cloud services used for podcasts had connectors of their own. [Descript's help center](https://help.descript.com/api-and-mcp/mcp.md?ref=implicator.ai) documented a link that lets outside assistants bring in files, hand changes to Descript's built-in agent and publish the results. A sample prompt there reads: "Import this video from my desktop into my 'Podcasts' folder and transcribe it." Riverside offered a beta connector to customers on its Grow plan and above, limited to material already recorded, with requests such as "Cut the filler words and long pauses from my latest recording and apply magic audio." Starting a recording still happens inside Riverside.

Around the vendors, volunteers kept publishing servers on GitHub. A REAPER server from TwelveTake Studios lists 180 tools. A Pro Tools server from Blue Elevator Productions carries a README dated Oct. 24 that calls it "Built for podcast post-production workflows."

In an interview, every spoken line has to stay tied to the moment in the recording where it was said, or a cut chosen on the page lands in the wrong place. Two voices recorded on separate tracks have to be cut together, or the host's question and the guest's answer slide apart by the length of each edit. And the original audio has to survive the agent's changes, so that a bad cut can be reversed and a disputed quote checked against the tape. The regions from the REAPER trial address the first condition. The other two depend on the workstation and on how the agent is allowed to work inside it.

II · REAPER

## REAPER, the editor agents can already drive

Every agent route into REAPER described here runs through ReaScript. Scripts written in Lua, EEL2 or Python work inside the application, and any built-in action, such as "Item: Split item under mouse cursor," can be called from them. The same scripts reach most of the programming interface that compiled extensions use, and each creates an undo point by default unless it calls `defer()`, the function that keeps a script running in the background. Lua and EEL2 are embedded; Python has to be installed separately. The current build, 7.84, is dated Oct. 28, 2026.

The pricing has two tiers. A $60 discounted license covers individuals for personal use, educational and non-profit organizations, and individuals or businesses using REAPER commercially whose yearly gross revenue does not exceed $20,000\. Commercial users above that line pay $225\. Neither is a subscription: either license is bought once and includes upgrades through version 8.99, and REAPER gives 60 days of evaluation with full functionality before payment. Prices exclude VAT and local taxes. A key can be installed on several computers as long as REAPER is used on only one at a time.

The agent layer comes from outside developers, and the servers are separate projects from REAPER itself. The [TwelveTake REAPER MCP](https://github.com/TwelveTake-Studios/reaper-mcp/blob/main/README.md?ref=implicator.ai), at version 1.9.4, advertises 180 tools. Inside REAPER it needs only a plain Lua script and an analyzer plug-in that come with it, so stock REAPER is enough; a separate Python server talks to the assistant. According to the README, every change the AI makes becomes a named step in REAPER's undo history, such as "MCP: delete track," so each press of Ctrl+Z reverses a single tool call. It measures loudness, loudness range and true peak for the master, a track or a single clip. Rendering and loudness measurement proceed at roughly real time, and the default wait is 600 seconds, so the README advises raising it for renders longer than ten minutes. Its developers ask for support through Buy Me a Coffee and Ko-fi.

`save_fx_preset` is marked "Not supported" and always returns an error. `render_region` is marked "Not implemented"; the list says to use `render_project` with the region's start and end instead. The README describes the full live suite as green "through REAPER 7.79." The current REAPER build is 7.84\. And the project states its purpose in its opening lines: letting AI assistants control REAPER "for mixing, mastering, MIDI composition and music production." Interviews are not on that list.

Two older extensions, ReaWhisper and ReaSpeech, already bring transcripts into REAPER for searching and marking passages. A review of their documentation found no evidence that deleting words from those transcripts cuts the recording underneath.

The engine is chosen in TypeWhisper: Parakeet works on the Mac, while ElevenLabs Scribe v2 and many others, reached through OpenRouter, are hosted services. In version 0.1.0, engines that returned timestamps produced a region per sentence or segment, and every engine wrote the full text into the item notes, a field attached to each clip in REAPER. No third-party relay sits in the middle, and nobody at the desk has to install or maintain its own copy of OpenAI's open-source speech model.

The Oct. 31 session ran three passes. The first used an aired radio feature of 13 clips, 7 of them voice edits and 6 English sound bites, sent to Scribe v2\. All six sound-bite transcripts matched the edited clips word for word. No regions appeared, because Scribe returned no timestamps, and the output landed in the notes.

The second pass sent the raw interview to Scribe, which refused it: "Payload too large. Maximum request body size is 256 MiB." Converted to mono at 16 kHz, the recording shrank to 206 MiB and went through. It came back as 16,213 words in a single item note, with nothing tying any of them to a point on the timeline.

The third pass ran the same recording through Parakeet and produced the 1,067 regions. A region in REAPER is a named span of the timeline, a labeled bracket over the waveform that can be jumped to, selected or rendered. The median bracket lasted 4.5 seconds and held 12 words. The ReaScript also did the bookkeeping that keeps those labels in place. Segment times arrive measured from the start of the source file, and the code maps them onto the item's position, trim offset and playrate, so each label is meant to sit over the speech it names wherever the item lives on the timeline. Parakeet and Scribe agreed on 91.8% of words, a measure of how closely two engines matched each other, not of accuracy against a checked transcript.

1,067 transcript regions from one 1:52:45 interview, Parakeet via TypeWhisper script 0.1.0 in REAPER 7.84, one session, Oct. 31, 2026 

The limits showed up in the detail. Most boundaries fell within about 0.3 seconds of the speech they mark, but 107 of the 1,067 regions lasted under a second, and those can miss the word entirely. The breaks fall at pauses, not at the ends of sentences. Left to detect the language on its own, Parakeet drifted into English partway through the conversation, which was not in English. Nothing marks who is speaking. A source file shared by several clips in a project is transcribed once per item, so a recording cut into pieces gets processed again for every piece. All of it went to Hillger as feedback the same day.

Version 0.2 came back on Nov. 1, and Hillger published the scripts on [GitHub](https://github.com/TypeWhisper/typewhisper-reaper?ref=implicator.ai) the same day under the MIT license. TypeWhisper now reads the recorder file directly, so the 1.95 GB WAV no longer needs converting. A source cut into several clips is transcribed once, regions follow sentence ends and speaker changes, and speaker labels, which require TypeWhisper Premium, mark who is talking. With TypeWhisper 1.8.2, Scribe also returns a timestamp for every word. Run on the unconverted interview on Nov. 1, it returned 16,121 words with their timings in 50 seconds. After the changes we contributed to version 0.2.1, all 771 regions on that interview carried a speaker label, 4% ended somewhere other than the end of a sentence, and 16 lasted under a second.

The cutting arrived the same day. "Cut quotes from transcript" takes passages marked with double square brackets in the transcript file and cuts them onto a track of their own. Each cut starts in the pause before the first word and ends in the pause after the last, with short fades and a "Cut" marker wherever two passages are joined. A third script, which we wrote and Hillger merged, treats a radio manuscript as the cut list. It finds each quote under a label line such as "Quote Miller:", cuts the quotes in manuscript order, reads an ellipsis inside a quote as a cut, and reports where the manuscript differs from what was said. In a test on the same interview, we marked three sound bites, one of them joined from two passages ten minutes apart. Each part held exactly the marked words, with no neighboring word cut in.

The timings still need checking against the audio. Measured at 150 sentence starts and 150 sentence ends in that interview, Scribe placed the ends of words closely, a median of 25 milliseconds before the speech stopped, but it placed starts late: the sound began a median of 140 milliseconds before the stated time, part of it probably breath. A cut made at the timestamp would often clip the first sound, so the scripts look for the cut point in the audio. Scribe sometimes puts a change of speaker in the middle of a sentence, and where both people talk at once, words go missing. The language setting is only a hint, and Parakeet can still switch languages in unclear passages. The scripts run on macOS only and are installed by hand.

None of the three scripts is an agent. Each does one fixed job when a person starts it. What they supply is the first condition an agent would need in an interview, words tied to their place in the recording, plus cut points checked against the sound.

III · INCUMBENTS

## Pro Tools and Audition, the incumbents

Pro Tools has carried a native Speech-to-Text engine since its 2025.6 update. The 2026.4 release, dated April 28, 2026, passes that transcription on to new files created by AudioSuite processing, Consolidate, Commit or Track Bounce, keeping word timing so nothing has to be transcribed again. The 2025.10 update had added splitting a clip by word, sentence or speaker in the Transcript window; 2026.4 extends that to timeline selections that cross several clips or tracks. Those 2026.4 Speech-to-Text enhancements are for Studio and Ultimate users. The Studio tier lists at $34.99 a month on Avid's site as of Oct. 31, excluding tax, and the $9.99 Artist tier is not among the tiers named. Avid offers a free 30-day trial.

Avid's own agentic work, as announced so far, is for a different product. A press release dated Aug. 27, 2026, ahead of IBC, previewed "a new agentic orchestration layer" for Media Composer, Avid's video editing application, along with deeper ties to Avid Content Core and a Google Gemini panel. The same application's 2026.8 update added tools to revert transcript edits and split segments, and PhraseFindAI Multicam Auto Cut, which places cuts on speech boundaries. Kenna Hilburn, Avid's chief product officer, said the approach meant "enabling more of the preparation and routine work to be automated while leaving editorial oversight and creative control firmly in editors' hands." Pro Tools does not appear in the announcement.

Third parties fill the space. Braus runs on the customer's own ChatGPT or Claude subscription, needs macOS 15 or later on Apple silicon, and is free during the beta. It reads the open session and acts on tracks, selections, files and the session itself, in the Mix and Edit windows. Forte's examples come from music work: rename four tracks Kick, Snare, Bass and Vox, import a folder of stems, mute the overheads. It can also report the sample rate, select time ranges, and open, save and close sessions. "That is the list today," Forte wrote. Cutting a clip is not on it. The company says the add-on takes a request at a time, and "It does not automate whole workflows, and it is not trying to." Forte has announced no later prices and says a free version will always exist.

Forte says that after each change, the add-on reads the session again to confirm the result and says when it could not check. If a series of track changes fails halfway, it reverses them, removing tracks it created and restoring names and mute, solo and record-arm states. Stopping it does not undo work already done. On data, Forte's post is specific. The add-on turns Pro Tools windows into text on the Mac and sends that text, with the request, to the connected AI. Screen images stay on the machine. Work on other Mac apps, DaVinci Resolve among them, "is early, no promises."

The community project from Blue Elevator Productions, protools-mcp, goes further into spoken word. It offers 25+ tools in seven groups, including search across the CSV that Pro Tools exports from Speech-to-Text, and Claude describes each write operation and confirms before executing it. Its writes cover selections, markers, mute and solo, consolidating a clip, saving, and AAF export. It needs Python 3.11 or later and, for the export dialog, macOS Accessibility permission. Underneath sits PTSL, the Pro Tools Scripting Library, which is Avid's official route into the application and listens on the local machine at port 31416.

Adobe's page for Audition 26.5, the September 2026 update, says only that it "includes improvements." The most recent feature Adobe lists for Audition this year is native support for Windows on ARM. On April 28, 2026, Adobe launched a Claude connector drawing on Photoshop, Illustrator, Firefly, Express, Premiere, Lightroom, InDesign and Stock. Audition is not on that list.

On April 15, 2026, Adobe unveiled Firefly AI Assistant, which it said would carry out multi-step work across Creative Cloud apps "including Firefly, Photoshop, Premiere, Lightroom, Express, Illustrator and more." Audition is not among the apps it names. Audition's June release, 26.3, removed the built-in CD authoring tools, and Adobe's community manager wrote that the release "does not introduce major workflow changes or new editing tools." The machine learning Audition does carry is older: Remix, which Adobe credits to its Sensei technology, refits music to a set length, and auto-ducking lowers music under dialogue. On April 8, a user posting as KarstenJensen asked in Adobe's feature-request forum for natural language editing commands, typed or spoken instructions such as "remove all background noise from 00:30 to 01:15," and to "have AI execute them on the timeline." By Oct. 31 the request was still open for voting. Audition costs US$22.99 a month on an annual plan billed monthly.

The only agent route into Audition found for this piece comes from outside Adobe. Brainferno's MCP server, which its README says is "not made or endorsed by Adobe," registers 12 Audition tools: application and document state, more than 600 menu commands, Favorites, which apply a saved effect chain hands-off, markers, transport, open, save and export. The README says every action is a single undo step and nothing leaves the computer unless shared mode is switched on. A sample request reads: "Open this WAV in Audition, apply my 'Podcast Voice' favorite, export it as MP3." It reaches Audition through ExtendScript, "via a CEP-only API learned by reflection." Under "Not yet," the same README lists "Audition multitrack writes." A two-voice interview on separate tracks lives in a multitrack session.

IV · CLOUD

## Descript and Resolve, the cloud and the video suite

Descript's connector signs in through the user's Descript login, so no API token is involved. During sign-in the user picks a single Drive, and the connection is scoped to that Drive alone; switching means logging out on the web and reconnecting. Using it from a team or enterprise workspace can require an admin to enable third-party connectors first. The help page lists the limits plainly, and the first reads: "No rendered file download without publishing." An assistant can publish to a Descript web link and fetch the result from there. A download without publishing has to be done by hand in the app. Transcripts can be exported as text, Markdown, HTML, RTF or SRT, but not DOCX, and job history disappears after 30 days.

Changes run through Underlord, Descript's built-in agent, which is still labeled beta. The help page says of it: "Underlord might overpromise, make incorrect assumptions, or follow you into workflows it's not equipped to complete." Each prompt is charged for what Descript calls the "agent brain," and charged again for any tool or effect it applies. Underlord is non-deterministic, so the same prompt can cost different amounts on different runs. It cannot edit in the background during a recording; the narration or script has to be in place first.

As listed at the end of October, Creator costs $24 per person per month billed annually, or $35 billed monthly, with 800 AI credits and 30 media hours a month. Hobbyist costs $16 a month billed annually, or $24 billed monthly, with 400 AI credits and 10 hours. Business, at $50 a person billed annually or $65 monthly, raises the allowance to 1,500 credits and 40 hours. Timeline export, the way out to an AAF for Pro Tools, an EDL for REAPER or SESX for Audition, requires Creator or a higher plan.

Vikram Singh bought the Creator plan in August 2025 and made Descript the main tool for a weekly interview podcast, shipping 47 episodes through May 2026\. He ran Underlord on six of them. "The pass-through rate (work I keep without major changes) is about 60%," Singh wrote in a review published May 10, 2026\. He had used the newest agent features for no more than four weeks at that point and called his view of them preliminary. He found transcription "noticeably worse on accents, technical jargon, and overlapping speakers." Working on the page also changed his habits at first: "Because it's so easy to delete sentences, you cut things you'd normally let breathe." His first three episodes felt rushed, by his account, because he had removed every pause longer than 0.4 seconds.

His most detailed warning concerns Studio Sound, Descript's noise-removal processing. "On about one in five clips, it adds a faint 'swimming pool' reverb to sibilants," he wrote. He caught it on episode 14 and has compared every pass against the original since, keeping the processed version about 75% of the time. For guests recorded over Zoom, he wrote, the compression artifacts confuse it, and he uses a separate plug-in instead. "Studio Sound has shipped one episode with audible artifacts that I should have caught." Underlord applies the effect on request; "Apply Studio Sound with 70% intensity" is among Descript's sample prompts.

Resolve's connection is built around a different job. The 21.1 update lets customers "analyze projects, organize media, adjust settings and batch render using everyday conversational language," and ask for highlight edits from long-form video, removal of unwanted clips and rendered deliverables. Grant Petty, Blackmagic's chief executive, said: "With AI assistant support, our customers can now ask Claude or ChatGPT to handle repetitive tasks such as create highlight edits, organizing media or batch rendering, which frees them up to focus on creative decisions." Blackmagic's feature sheet lists AI assistants for the paid edition on desktop only, not on iPads, and notes that the feature "Requires an MCP-compatible AI assistant."

Resolve Studio lists at $295\. A sale announced Oct. 20 cuts the price to $236 until Nov. 20, 2026\. The examples Blackmagic gives, highlight reels and rendered deliverables, are written for shows that also publish video.

V · POLICY

## Ardour and the accessibility argument

Ardour's server stays off until someone turns it on. In 9.5 it sits under Edit > Preferences > Control Surfaces as a checkbox labeled "MCP HTTP Server (experimental)." The 9.5 announcement calls it community-contributed and meant for "a locally running large language model." It names a practical use, better accessibility for people with visual impairments, and adds that even commercial workstations fall short there. It also draws two lines: "this implementation does not extend to generating audio with LLMs. Nor do we accept LLM-generated patches".

The policy behind it was argued in public. A forum thread titled "What is Ardour's Generative AI policy?" ran from April 30 to July 20, 2026, and drew 96 replies. It opened with a user asking why such a feature was coming at all, after an earlier forum answer had put the role of generative AI "in anything related to Ardour" at "zero."

Paul Davis, Ardour's founder and lead developer, replied on April 30 that the policy rested on the project's own interests, "despite the two main developers' severe concerns with LLMs in general." The worry he named was reviewers being swamped by machine-written submissions, and code "that has subtle but deep misconceptions" slipping into the codebase. "So at this time, 'LLM-related functionality' is acceptable."

On May 3 a user posting as cloudskat3r, who had used Ardour for about a year and supported it financially since switching, asked for the opposite. "I earnestly and respectfully ask you to consider banning anything to do with 'generative' AI outright." The reasons, in the post's own words, included "it's environmental impact, it's effect on the price of hardware," and "its ties to the war industry." The same post described pride in an audio workflow "entirely FOSS and AI free." Ten days later, writing as someone with a disability that makes visual art harder to create, cloudskat3r rejected the accessibility case: "I doubt anything AI related has to be involved."

Davis posted in reply that the alternative would be hard. "If you wanted to build an interactive front-end that allows a person to take control of Ardour from speech or text inputs, it would be a huge task if you chose to do it any other way than the LLM way." Robin Gareus, the other main Ardour developer, answered the same day the suggestion that such accessibility functions could be found another way. "Highly unlikely and not for a lack of trying. We failed to do that for 20+ years."

> We failed to do that for 20+ years.
> 
> Robin Gareus, Ardour developer, May 13, 2026 

Both developers had also set out limits. Gareus wrote on May 12 that "Free as in freedom software does not discriminate against any use," and added: "There are also no plans to directly integrate any genAI, though interfaces for 3rd parties to use at their digression are welcome." Davis, the same day, said the project was "not seeking to make it easier for AI tools to use Ardour than it already is (which is actually pretty easy)." Pressed on the wider ethics, he wrote the next day, "I'd rather work on bugs and features."

Ardour costs what the buyer decides. A single payment under US$45 buys the current version and its updates; $45 or more adds the next major version and access to nightly builds. Subscriptions run $1, $4, $10 or $50 a month. The free demo periodically goes silent after 10 minutes.

VI · PILOT

## A proposed test before the first real interview

What follows is a proposed pilot record, not a finding. We have not run it, and it sets no pass marks beyond the delivery spec a show already works to.

**Starting conditions.** Work on a duplicate of the project, never the original, and keep the recorder files outside the folder the agent works in. Choose a 12-minute excerpt of a two-voice conversation, with each voice recorded on its own track. Before the first request, write down the build of the workstation and of the agent connection, for example REAPER 7.84 with TwelveTake 1.9.4, or Pro Tools 2026.4 with the Braus beta.

**Steps.** Ask the agent for three bounded edits, in sequence: cut a named answer, add markers where the topic changes, render the excerpt. Phrase each request as a producer would on deadline, naming the answer by what the guest says rather than by timecode. After each, confirm that the change can be reversed, ideally in a single step of the undo history.

**Expected result and the check.** Every changed item keeps its position and source offset except where the request asked for a change. Both voice tracks are cut at the same points. Someone listens at every changed boundary, on both voices, before anything else happens. Loudness and true peak are measured on the rendered file and compared with the delivery spec the broadcaster or podcast host sets. Where a render can include or skip the project's own processing, as TwelveTake's `render_project` can, note which was used.

**Record.** The transcription engine and its settings, including whether the language was detected automatically, since Parakeet drifted into English in the trial when left to choose. Disagreements between engines, if two were compared. Missed words in very short segments, the failure seen in pieces under a second. Whether speakers were labeled. Review time, from the first request to sign-off. Credits or media minutes used where Descript is involved, keeping in mind that the same prompt can cost different amounts. What left the machine: Braus sends a text description of the session to the connected AI, while TypeWhisper can keep the work on the Mac with Parakeet or send the recording to a hosted engine such as Scribe.

**Decision owner.** The producer who signs off the episode decides whether the agent may touch the next real interview.

VII · COMPARISON

## Which editor fits which producer

| Six editors, one question: how far can an agent reach in? ● marks a documented capability; three is the most found. ○ marks one absent or not documented. Ratings are our reading of vendor and project documentation as of Oct. 31, 2026, TypeWhisper as of Nov. 8, not a benchmark. EditorAgent controlEdit by transcriptBuilt-in AIPrice (US)The catch REAPER\+ TypeWhisper scripts, open source●●○community MCP server, 180 tools●●○add-on: cut sound bites from a marked transcript or manuscript○○○none from Cockos$60 one-timerevenue up to $20,000; $225 commercial; upgrades through 8.99Server is built for music production and tested through 7.79; TypeWhisper cuts are scripted, not agent-driven, macOS only Pro Tools Studio●●○third parties via Avid's scripting SDK●●○native speech-to-text, select by word or speaker●○○transcription engine$34.99/moexcl. taxNo agent from Avid; Braus is a free beta that sends a text description of the session to your AI Descript●●●vendor MCP●●●delete text, cut audio●●●Underlord, Studio Sound$24/moCreator, billed yearly; $35 monthly; 800 AI creditsNo rendered file through MCP without publishing; the same prompt can cost different credits DaVinci Resolve Studio●●●vendor AI-assistant support, 21.1●●●text-based editing●●●Neural Engine$295 one-time$236 until Nov. 20, 2026Assistants work in Studio on the desktop only; built around video Ardour 9.5●●●shipped MCP server, opt-in○○○none○○○no generative AI, by policyPay what you want$45+ adds next major versionMeant for a locally running model; community-contributed server Adobe Audition●○○third-party server, 12 tools○○○none●○○Remix, auto-ducking$22.99/moannual plan, billed monthlyNot in Adobe's Claude connector; the third-party server lists multitrack writes as "Not yet" Agent control: three dots for a connection the vendor ships, two for a broad server from a third party or the community, one for a narrow one. Edit by transcript: three dots when deleting text cuts the audio, two when marked text is cut into sound bites. Sources are linked in the article. |
| -------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- |

REAPER's license is paid once: $60 for an individual or a business under $20,000 in yearly gross revenue, $225 otherwise, upgrades through 8.99 and 60 days to evaluate. A Lua script and a Python server are the whole install. That reach comes from community code. TwelveTake's server was written for music production, and its live suite is listed as green through 7.79\. TypeWhisper's open-source scripts lay an interview's words on the timeline and cut sound bites from a marked transcript or a manuscript, tested so far only by us, on one Mac, as contributors to the project.

A station that already runs Pro Tools gets native Speech-to-Text on its two upper tiers, the cheaper at $34.99 a month, including splits by word, sentence or speaker. The $9.99 Artist tier is not among them. Avid's agent features, as shown in August, belong to its video editing software. Agent control in Pro Tools comes only from third parties so far: Braus, free in beta and limited to housekeeping on tracks and files, and protools-mcp, a community project that searches the transcript and asks before every write.

Descript's route runs through the cloud, on conditions. Its connector reaches a single Drive, a finished mix cannot be pulled through it without publishing, and every agent request draws on metered credits at a cost that can vary from attempt to attempt. Exports to REAPER, Pro Tools or Audition start with the Creator plan, $24 a person a month billed annually. Singh, in an account he called preliminary, put the share of Underlord's work he kept without major changes at about 60%.

Resolve Studio, whose assistant examples are written for video, costs $236 until Nov. 20 against $295 at full price, with assistant support from the vendor on desktop only.

Ardour offers an open-source, opt-in route for a local model, at whatever price the buyer chooses, from developers who have said they are not seeking to make AI use easier.

Audition, at US$22.99 a month, has no agent route from Adobe as of Oct. 31, 2026, while REAPER's $60 license is paid once. The third-party route found, Brainferno's 12 tools for files, menus, markers and export, cannot yet write to multitrack projects.

VIII · SIGNALS

## What to watch

**TypeWhisper** 

transcription of each speaker's recorded channel on its own, which Hillger has named as the next step for interviews and which would settle who said what when voices overlap, and watch folders in the TypeWhisper app that write the transcript notes the REAPER scripts read. No date has been given for either.

**TwelveTake** 

a version after 1.9.4 whose live suite covers REAPER 7.84, the current build. The README lists the suite as green through 7.79, and no date has been given for the next check.

**Forte** 

prices for Braus once the beta ends, which Forte has not announced beyond promising a permanent free version, and its work on Resolve, which it calls "early, no promises." No timeline has been published for either.

**Avid and Adobe** 

whether Avid brings the agent features it showed for Media Composer at IBC 2026 to Pro Tools, and whether Adobe adds Audition to that Claude connector. Neither company has given a date.

**Blackmagic** 

the $236 sale price for the paid Resolve edition ends Nov. 20, 2026, against a regular price of $295.

Robin Gareus, in the post that answered the accessibility point, also described the job he had in mind. "I also like the fact that LLMs can be used as studio-assistant. Your new intern can set up routing, and do basic session tasks…"

Frequently Asked Questions

Can an AI agent edit an interview in REAPER today?

Through community servers such as TwelveTake's REAPER MCP, which lists 180 tools and, according to its README, records each change as a named undo step. It was written for music production and marks region rendering as not implemented. TypeWhisper's open-source scripts place transcript regions on the timeline and cut sound bites from passages marked in the transcript or quoted in a manuscript, but each runs one fixed job when started and none is an agent.

Does Pro Tools have an AI assistant?

Not from Avid. Pro Tools has a native Speech-to-Text engine, and the 2026.4 release extended splitting by word, sentence or speaker to timeline selections for Studio and Ultimate users. Agent control comes from third parties: Forte's Braus beta, which runs on your own ChatGPT or Claude subscription, and the community protools-mcp server, which confirms before each write.

Why doesn't Adobe Audition work with Claude?

Adobe has not connected it. The Claude connector Adobe launched on April 28, 2026, draws on Photoshop, Premiere and six other products but not Audition, and the app list in its Firefly AI Assistant announcement does not name Audition either. The only route found is Brainferno's third-party MCP server with 12 Audition tools, whose README lists multitrack writes as not yet supported.

What is the catch with Descript's connector?

It is scoped to a single Drive, cannot deliver a rendered file without publishing the project, and can use media minutes and AI credits. Descript warns that Underlord "might overpromise," and says the same prompt can cost different amounts on different runs.

What does a REAPER license cost?

$60 for individuals and for businesses whose yearly gross revenue does not exceed $20,000, and $225 for commercial users above that. Either license is bought once and includes upgrades through version 8.99, after a 60-day evaluation with full functionality.

AI-generated summary, reviewed by an editor. [More on our AI guidelines](https://www.implicator.ai/about/).

### Sources & Further Reading

#### Primary sources

- [Ardour 9.5 release notes](https://impli.me/MXAvMv?ref=implicator.ai), Ardour, May 19, 2026
- [What is Ardour's Generative AI policy? (forum thread)](https://impli.me/qgOtWj?ref=implicator.ai), Ardour Discourse, April 29 to July 20, 2026
- [TwelveTake REAPER MCP README](https://impli.me/YnVPjh?ref=implicator.ai), GitHub, version 1.9.4
- [Purchase a REAPER license](https://impli.me/mAIsuY?ref=implicator.ai), Cockos
- [Pro Tools 2026.4: what's new](https://impli.me/BlKQDf?ref=implicator.ai), Avid, April 28, 2026
- [Braus for Pro Tools](https://impli.me/FP2w8a?ref=implicator.ai), Forte, Oct. 27, 2026
- [DaVinci Resolve 21.1 announcement](https://impli.me/x0LQ8r?ref=implicator.ai), Blackmagic Design, Sept. 8, 2026
- [Descript MCP overview](https://impli.me/FSWGLs?ref=implicator.ai), Descript Help Center
- [Adobe for creativity connector for Claude](https://impli.me/tza1N1?ref=implicator.ai), Adobe, April 28, 2026

#### Community projects and reviews

- [TypeWhisper for REAPER](https://github.com/TypeWhisper/typewhisper-reaper?ref=implicator.ai), Marco Hillger, GitHub, MIT license
- [protools-mcp README](https://impli.me/FwMfgE?ref=implicator.ai), Blue Elevator Productions, GitHub
- [Brainferno MCP Bridge README](https://impli.me/tX1FIm?ref=implicator.ai), GitHub
- [Descript Review 2026: I Edited 47 Episodes With It](https://impli.me/tnYxEH?ref=implicator.ai), Vikram Singh, May 10, 2026

## Related stories

[Claude Code and DaVinci Resolve: 5 Best Editing TasksClaude Code can operate inside DaVinci Resolve Studio. The useful question is which jobs it handles well, and where the editor still has to intervene.Implicator.ai![](https://www.implicator.ai/content/images/2026/08/2026-08-19-11.27.10-resolve_editing_tasks@2x.webp)](https://www.implicator.ai/claude-code-davinci-resolve-five-editing-tasks/)