TIMEBOLT JUST GOT A BRAIN.

Edit video with
ChatGPT or Claude.

TimeBolt’s MCP server connects ChatGPT or Claude to your video transcript and timeline. Make cuts you can see, hear, and change.

01

Clean it up.

Tighten this scripted recording. Remove abandoned takes, retries, repetition, and unnecessary filler words. Keep the strongest complete take of each point and every essential step, warning, and qualification. Preserve useful pauses for on-screen actions; flag anything needing visual review. Add occasional subtle 1.15× centered punch-ins where they help without cropping important content. Add YouTube chapters timed to the finished edit. Keep selected passages in source order and preserve meaning. Use the supplied transcript and word timestamps. Keep the original framing and playback speed. Apply the edit in TimeBolt for review. After approval, export captions retimed to the finished edit from the supplied SRT; omit removed speech and use word timestamps to split cues at cuts.

About this test

The prompt at left is a reusable template; these exports were produced during iterative testing.

02

Make it shorter.

Find the strongest standalone short in this scripted recording. Aim for 30–60 seconds, with a compelling existing opening, one complete idea, and a satisfying ending. Preserve context and choose the strongest complete takes. Remove retries and repetition. Preserve pauses needed to understand on-screen actions. Suggest a title and report the selected source ranges. Keep selected passages in source order and preserve meaning. Use the supplied transcript and word timestamps. Keep the original framing and playback speed. Apply the edit in TimeBolt for review. After approval, export captions retimed to the finished edit from the supplied SRT; omit removed speech and use word timestamps to split cues at cuts.

About this test

The prompt at left is a reusable template; these exports were produced during iterative testing.

03

Multiple cameras.
One timeline.

Bring every
track together.

Your cameras. Your master audio.
One synchronized edit.

CAM 1CAM 2CAM 3MASTER AUDIO

You direct.
AI makes the cuts.

Use with ChatGPT or Claude

Edit this VSL 6 multicam: best takes only, no filler or dead air. Switch cameras to fit the content, keep VSL 6 as the only audio, and add YouTube chapters.

Keep the
final say.

Change an angle. Restore a cut.
Fine-tune before you render.

Verify and fine-tuneTimeBolt · actual footage

04

Connect your AI.

Download TimeBolt ↗

In TimeBolt, open Local MCP → Connect to ChatGPT Desktop… and follow the confirmation.

Connection help

Use the desktop assistant named in TimeBolt’s Local MCP menu. Follow the connection confirmation. For Claude, fully quit and reopen the app, then review Settings → Connectors → TimeBolt. Open your media and prepare its transcript with UMCheck or a supported transcript import before asking for an edit.

Local rendering does not mean your AI conversation is offline.

Your next move.

Render your video in TimeBolt. Export an editable timeline to Final Cut Pro, Adobe Premiere Pro, or DaVinci Resolve.

TimeBolt logoMP4

Render your finished video.

Review your cuts, choose your output settings in TimeBolt, and render on your computer.

  1. Review the edit
  2. Choose output settings
  3. Render locally
Get TimeBolt ↗

A few useful answers.

AI video editing with TimeBolt.

What is the TimeBolt MCP connector?

The TimeBolt MCP connector lets ChatGPT or Claude work with your transcript and editing timeline. Ask for cuts in plain language, then review and refine the edit in TimeBolt before rendering or exporting.

Can I use ChatGPT and Claude for video editing?

Yes. Connect either supported desktop assistant through TimeBolt’s Local MCP menu. Open your media, prepare its transcript with UMCheck, and give your AI an editing prompt.

Can AI edit a multicam recording in TimeBolt?

Yes. TimeBolt supports synchronized audio and video tracks, camera switching, and timeline editing through MCP. Specify your master audio track and camera instructions, then check the cuts and switches in the multitrack editor.

Can I change the cuts AI makes?

Yes. Review the applied edit in TimeBolt, restore a passage, adjust a cut, or change a camera angle before you render. You keep the final say.

Can I export to Final Cut Pro, Premiere Pro, or DaVinci Resolve?

Yes. Export FCPXML for Final Cut Pro or XML for Adobe Premiere Pro and DaVinci Resolve. These are editable timeline exports; rendering a finished video is a separate option. Saving and timeline export require a paid TimeBolt account.

Does local editing mean my AI conversation is offline?

No. TimeBolt renders locally, but your connected AI may process transcript text and tool results through its own service. Review your assistant’s permissions and data settings.

Do I need AI to use TimeBolt?

No. TimeBolt is a standalone desktop video editor. Connecting ChatGPT or Claude is optional.

UMCheck transcription ↗TimeBolt features ↗MCP tool reference ↗Known behaviors ↗

Explore the earlier edit test +

See what went.
Bring back what matters.

An AI edit is a set of decisions. Open a removed passage, read the context, and decide whether it belongs in your story.

OFFICE HOURS · ACTUAL TEST03:10.51–03:44.95
PASSAGE REMOVED1:51.34 kept

“…could be the first bring your own model to TimeBolt specifically Umcheck.”

“I think that particular element of reducing token cost…”

03:03Word timings · not audio waveform03:58
Would your audience
need this explanation?

Actual transcript & cut timings. Browser simulation; your project stays unchanged.

↶A cut you can reverse

The recorded edit removed 34.44 seconds explaining transcript import. Keeping that context changes the timeline from 1:51.34 to 2:25.78.

  • Open the passage to read what was removed.
  • Restore it without changing the other cuts.
  • Review the meaning and sound in TimeBolt.
30mOne recording. A real applied edit.

A founder-directed Codex session condensed a 30-minute recording into 1:51 of kept timeline regions. The test used a more aggressive prompt than the suggested instruction above.

Rendering remained pending. This measures kept duration, not processing speed.

±What about cut-boundary correction?

Automatic correction has not been demonstrated for this MCP workflow. A reported FCPXML test matched submitted timestamps within one frame.

This example shows restoring a passage, not automatic acoustic correction.

Open the test record 18 September 2026

An OpenAI assistant in Codex desktop opened a local unscripted Zoom recording, read a prepared transcript and the existing timeline, then applied cuts. Transcript preparation required user action. MP4 and Turbo +12% settings were not exposed through MCP, so rendering remained pending.

“What I would like you to do is to remove anything that's not absolutely and vitally important to the conversation and the topic being discussed, including filler words, banter, stuff like that, dead air, obviously. And then I would like you to make two shorts based on the best content and render them as an MP4 and use Turbo to speed that up by 12%.”

A read-only check on September 18, 2026 found 75 keep/remove regions, 1,800.885375 seconds of source media, and 111.3355 seconds of kept regions before speed adjustments. This verifies timeline duration, not a rendered file or processing speed.

Model and TimeBolt build: not established in the reviewed record. A single demonstration, not a controlled benchmark or promise of identical results.

Two short selections: approximately 29 and 24 seconds. Identified, not rendered or vertically reframed.

This interactive reconstruction uses the recorded transcript and cut intervals. Its suggested prompt and restore control are demonstrations, not additional actions from the original session.

REAL PRODUCT DEMONSTRATION

Claude + TimeBolt

Separate from the Codex test explored on this page. One founder-run review-and-repair comparison. Both workflows required judgment.

Watch on YouTube ↗
MCP tool reference+

Reference reviewed September 18, 2026. Capabilities can change with your installed TimeBolt version.

Reading

ToolInputDocumented purpose
get_file_path{}Read the open media path
get_duration{}Read duration; observed source duration, not kept-output duration
get_timeline{}Read existing timeline before editing
get_umcheck_transcription{}Read prepared word-level transcript; if empty, prepare it in TimeBolt or supply a transcript
get_predefined_filler_words_and_phrases{}Read predefined comma-separated filler list
get_render_queue{}Read remaining rendering tasks

Opening and writing

open_file_in_timeline: input { "filePath": "/absolute/path/to/video.mp4" }. Opens new media; not a way to import transcription JSON.

apply_timeline outer input schema:

{
  "type": "object",
  "properties": { "timeline": { "type": "string" } },
  "required": ["timeline"]
}

The timeline string contains a serialized JSON array, not a raw array argument. The server description gives these inner constraints:

FieldContract supplied in description
start, durationSource interval coordinates; inspected values use seconds
typekeep, remove, fastforward
punchedAt least 1, strictly less than 5
punchedPositioncenter, top-left, top-right, bottom-left, bottom-right, top, bottom, left, right
operationField is named, but its allowed values are not specified in the exposed description; inspected keep/remove regions mirror type
markedBoolean
titleString; identifies the start/title of a chapter or topic

Read and modify the existing timeline; do not discard prior edits. The description requires no gaps, with uncovered gaps explicitly represented as remove regions. The description does not provide a formal inner JSON Schema, a maximum payload size, overlap policy, fast-forward speed semantics, or error/return schemas. Do not pretend those are established.

Illustrative whole timeline for a fictional 3-second source, not an instruction to overwrite the current project:

{
  "timeline": "[{\"start\":0,\"duration\":1,\"type\":\"keep\",\"punched\":1,\"punchedPosition\":\"center\",\"operation\":\"keep\",\"marked\":false,\"title\":\"\"},{\"start\":1,\"duration\":2,\"type\":\"remove\",\"punched\":1,\"punchedPosition\":\"center\",\"operation\":\"remove\",\"marked\":false,\"title\":\"\"}]"
}

Rendering and export

All tools below take an empty object in the documented usage.

  • add_to_render_queue: queue the current timeline.
  • start_render: start all queued tasks; check the queue before invoking it.
  • generate_fcpxml: generate Final Cut Pro XML; save it and return its path.
  • generate_fcpxml_multicam: generate multicam FCPXML.
  • generate_xml: generate XML for Premiere Pro / DaVinci Resolve.
  • generate_xml_sequence: generate nested-sequence XML for Premiere Pro.
  • generate_srt: generate a caption file from timeline/transcription.

The exposed tools do not configure render format, Turbo, vertical crop or aspect ratio, start UMCheck, or import an existing UMCheck JSON by path. Exporting FCPXML does not itself prove boundary correction.

Configuration

Current supported customer instructions: Claude Desktop setup. The application provides a connection command. Manual configuration JSON is not established by this tool listing. Add platform-specific, tested JSON and a Copy action only after the connector maintainer supplies it. Never invent a localhost port or executable path.

Proposed adjacent documentation: /docs/mcp/known-behaviors. Both paths are planned, not currently verified public pages.

Known behaviors →

Known behaviors+

Reference reviewed September 18, 2026. Capabilities can change with your installed TimeBolt version.

get_timeline returns type: "original"

Reported in a connector test. The current write description allows keep, remove and fastforward, not original. Do not round-trip an original value unchanged into apply_timeline. Inspect its intended meaning and ask the maintainer to document normalization; do not blindly assume every original region means keep. Preserve source coverage and other cut properties. The current inspected timeline contains keep/remove regions, so the original-read state was not reproduced here.

PayloadTooLargeError on apply_timeline

Reported in a connector test; exact byte limit, responsible layer and affected build are not verified. Avoid retrying the same oversized body repeatedly. Minify JSON and remove redundant serialization whitespace. Count the UTF-8 bytes of the entire outer request, including string escaping; the transport may add further overhead. Preserve the complete timeline. Do not split a replacement timeline into separate applies unless an append/patch contract is documented. Merging adjacent regions requires identical metadata and semantics and may erase intentional editing boundaries; do not do it silently. Escalate with the measured request size, build/client and exact error, not private media or transcripts.

get_duration returns source length

User-reported and supported by the read-only check in this task: the tool returned 1800.885375 seconds with a heavily reduced keep/remove timeline. Its short description calls this total timeline duration, so documentation must clarify the source-time basis. For a pure keep/remove edit without speed changes, sum kept durations to estimate output length. Do not use that sum as rendered duration when Turbo, fast-forward or other timing adjustments apply. Rendered media is the authority for final output duration.

apply_timeline returns nothing on success

Reported in the user's connector test; no explicit success-response contract is exposed. An empty response alone does not verify success. Read the timeline back and compare intended intervals/properties before claiming completion. Do not repeat a write simply because the response is empty; first resolve the state through read-back. Check for a transport/tool error separately.

FCPXML boundaries match submitted timestamps

User reported exported boundaries within one frame of the submitted values. This is consistent with execution of the submitted intervals; it does not establish acoustic correction. A future correction claim requires submitted/applied/exported coordinates plus audible evidence on the same source.

What to include in a reproducible report

TimeBolt build, operating system, client and connector version, tool name, date, synthetic/minimal request, measured request bytes where relevant, response/error, and read-back comparison. Exact model if known; otherwise say unrecorded. Do not replace an unknown version or limit with a guess.

Tool reference →