Ostia is a fast video editor for Mac, written in Rust on Apple's own frameworks. You can edit by hand on the timeline, or let an AI app such as Claude, ChatGPT or Cursor drive it.

On an Apple M5 Pro it opened a 1,058-clip project in 2.95 ms (in-app median), played a 7-minute 4K60 timeline with no dropped frames in one full pass, and exported 4K60 HEVC at 2.16× realtime on a quiet machine.

Which AI apps can edit video with Ostia?

Any app that speaks MCP (Model Context Protocol). The ostia connect command sets up Claude Code, Claude Desktop, the ChatGPT desktop app and Codex, Cursor, Zed and Gemini CLI. Claude Code and Codex are tested end to end. Inside the app, ⌘L opens Claude beside the timeline.

Does my footage leave my Mac?

Transcription, voice cleanup, retouching and music run on your Mac. An AI app gets only what its tools return: transcripts, and frames it asks for. Optional stock footage downloads are the only other network use.

Can Ostia open CapCut projects?

Yes. Ostia imports a CapCut draft as a new Ostia video and leaves the CapCut files untouched.

What do I need to run Ostia?

A Mac with Apple silicon and macOS 14 or later. Claude features need Claude Code. Auto edit and transcripts need a 1.6 GB speech model, downloaded separately.

Footage docks here.
Films set sail.

Ostia is the port for your video: a fast Mac video editor written in Rust and built to be driven by Claude, ChatGPT, Cursor and more. Raw takes come ashore, get cut, captioned and loaded, and leave as finished film. 4K60 departs at 2.16× realtime.

Download for Mac Why it's fast

Apple silicon · macOS 14 or later

harbor clock3:22

film loaded7:17

2.16×realtime · 4K60 HEVC

fig. 1 — at seaheading —range —this page — fps

Fast, for reasons you can check.

launch to window 302ms

1 s target

median of 15 launches

open 1,058 clips 2.95ms

one 120 Hz frame

in-app median · tick: slowest, 4.94 ms

frames dropped 0of 26,211

one full pass

7 minutes of 4K60

4K60 export 2.16× realtime

film length

quiet machine, two runs

first open of a set: 207–212 ms request to reply

fig. 2

Nine reasons

One native binary 40.3MB editor, command line, agent server

Frames never copied ~4ms GPU per frame · prototype, Nov 8

one export

The CPU mostly waits 92–98% of export is the hardware encoder

before the fix 116 ms

Decoders ready before cuts 14.42ms longest decode wait

30 ms target

Proxies for scrubbing 3.3–15.2ms median

one frametwo frames

Edits inside one frame 5.8–7.9ms p95 · slowest 9.75 ms

200 ms target

Audio follows the cut 3.79ms p95 · headless

CPU 145.9 s

Speech on the GPU 29.7s 18-minute recording, Nov 9

other framework 386 MB

Interface on the GPU 289MB test, Nov 9

Measurement notes

Timings stop when a frame finishes drawing, not on screen.

Composite: a 60-second, 3-layer 4K clip on the prototype engine.

Scrub: heaviest 30-second stretch, 3 runs.

Exact frames while dragging take 32–47 ms.

Speech took 50.6 s when the GPU was shared.

The slowest edit, 9.75 ms, took longer than one 8.33 ms frame.

fig. 3

Same cargo, two ships

automatic edit, one 5-minute episode

Python pipelineone cold run, heavier load

transcriberetakesreview

17:53

Ostiaearlier build, median of 3

2:21

05:0010:0015:00

7.58×faster 15:31 sooner

100× speed

Stage by stage

Stage	Python	Ostia	Ratio
Audio read	3.71 s	4.05 s	0.92×
Sync	0.43 s	0.28 s	1.5×
Transcription	206.26 s	42.65 s	4.8×
Take re-transcription	467.95 s	52.25 s	9.0×
Align, cut, timing	31.5 s	0.09 s	—
Captions, timeline, save	1.97 s	0.03 s	—
Voice cleanup	12.91 s	1.22 s	10.6×
Review	362.28 s	40.91 s	8.9×
Total, measured	1,072.86 s	141.47 s	7.58×

Lanes are drawn to the measured totals.

The third segment is the rest of the measured total.

Stage rows were timed separately.

The Python rows add up to 1,087.01 s.

Audio read was not faster in Ostia.

Align and save: the source log gives no ratio.

On an 18-minute recording: 910.98 s against 146.69 s, 6.21×.

fig. 4

26,211 frames. None lost at sea.

dropped 0gaps over 25 ms 0

0:007:17

one full pass · two 4K layers · one dot per frame

fig. 5

Built to be driven by Claude, ChatGPT, Cursor and more.

Claude
ChatGPT
Cursor
Zed
Gemini

one MCP server, ostia connect sets each up · tested end to end with Claude Code and Codex

Claude · Ostia⌘L

youTighten the pause after the intro.

claudetranscriptwords, timings

edit ripple_cut1 undo step

check_overlaps0 overlaps

export 4K60started

staged session · real tool names

fig. 6

54tools over MCP

library 10

plan 4

look 9

timeline 12

auto edit 4

finish 8

checks 4

export 3

called in fig. 6grouped by us for this page

fig. 7

Claude beside the timeline⌘L
It knows where you areprojectvideoplayheadselectionunsaved
One undo step per change
Heavy work runs on your Mac
Rules kept per project

What Claude can see

Claude gets what its tools return: transcripts, and frames it asks for.
Transcription, voice cleanup, retouching and music run on your Mac.
Optional stock footage downloads are the only other network use.
The panel runs your installed Claude Code, signed in to your account.
The edit log records whether you or Claude made each change.

Plan to upload-ready in one port.

plan
autosave 2 s
lay out
387 clips in 0.27 s
shoot
sync 0.5 s
edit
229 lines in 3:38
finish
intro captions 45 smusic 5.7 s
export
4K60 HEVC, 10-bit
upload-ready
mp4

15:51

raw recording to checked file

run

15:51 raw

10:38 film

7:17

one scripted run · sync, auto edit: 3 benchmark runs

fig. 8

What the clock includes

One scripted end-to-end run, 951 s, with captures, checks and waiting.
Inside it the edit took 3:52 and the export ran at 1.72×.
Writing, shooting and human review are not included.
Stops are evenly spaced; they are not on a time scale.
Captions behind the speaker were made for the intro only.
Skin retouching was measured only on a short test clip.

Three exports of one film, each under 7:17.

film

7:17 quiet machine

3:22 scripted run

4:14 while screen mirroring

5:40

0:007:17

100× speed

4K60 HEVC 10-bit · quiet run measured twice

fig. 9

os·ti·a /ˈɒs.ti.ə/

ostium Latin · a mouth, a door

plural: ostia

fig. 10

Rust on Apple's own frameworks.

122,250lines of Rust

26crates

0Python at runtime

GPUI interface whisper.cpp speech MCP server agents

Metalcompose

VideoToolboxdecode, encode

CoreTexttype

Apple silicon

v0.1.0 · counted Nov 13

fig. 11

Questions.

What is Ostia?: Ostia is a fast video editor for Mac, written in Rust on Apple's own frameworks. You can edit by hand on the timeline, or let an AI app such as Claude, ChatGPT or Cursor drive it.
How fast is Ostia?: On an Apple M5 Pro it opened a 1,058-clip project in 2.95 ms (in-app median), played a 7-minute 4K60 timeline with no dropped frames in one full pass, and exported 4K60 HEVC at 2.16× realtime on a quiet machine.
Which AI apps can edit video with Ostia?: Any app that speaks MCP (Model Context Protocol). The ostia connect command sets up Claude Code, Claude Desktop, the ChatGPT desktop app and Codex, Cursor, Zed and Gemini CLI. Claude Code and Codex are tested end to end. Inside the app, ⌘L opens Claude beside the timeline.
Does my footage leave my Mac?: Transcription, voice cleanup, retouching and music run on your Mac. An AI app gets only what its tools return: transcripts, and frames it asks for. Optional stock footage downloads are the only other network use.
Can Ostia open CapCut projects?: Yes. Ostia imports a CapCut draft as a new Ostia video and leaves the CapCut files untouched.
What do I need to run Ostia?: A Mac with Apple silicon and macOS 14 or later. Claude features need Claude Code. Auto edit and transcripts need a 1.6 GB speech model, downloaded separately.

Set sail.

Download for Mac

v0.1.0 · 19.5 MB · Apple silicon · macOS 14 or later

Claude features need Claude Code · speech model 1.6 GB, not included

First open

macOS stops the first open of this early build.

Allow it in System Settings, under Privacy & Security, then open it again.

Auto edit and transcripts need two speech model files:

ggml-large-v3-turbo.bin · ggml-silero-v5.1.2.bin

Put them in ~/Library/Application Support/Lindy Studio/models/

Footage docks here.Films set sail.