Releases: Zyvro/Zyvro-desktop
Release list
Zyvro Studio v0.1.0-alpha.7
Zyvro Studio is a desktop workspace for Zyvro workflows: a file tree, an editor,
a terminal, source control, and an agent panel, with the graph editor opening in
a tab beside the code it acts on.
This is an alpha. It is built from three public repositories, it is not
signed, and the parts that touch a store are new. Read the two warnings below
before the first launch.
Why it exists
A ChatGPT or Claude subscription is not an API credential, and no hosted server
can use one. But the claude and codex command line tools on your own machine
are already signed in, so Zyvro Studio runs them as text providers. Your
subscription drives a workflow, and no credential ever reaches Zyvro.
Everything runs locally. Workflows live as files under .zyvro/ in the folder
you open, so they version alongside your code. The only traffic that leaves your
machine is the model call itself, and it goes straight to the provider.
First launch
macOS. Take the .dmg for your machine — arm64 for Apple Silicon, the
other for Intel — open it and drag the app to Applications. A .zip of the same
build is there too, for anyone who would rather not mount an image.
alpha.2 shipped without a .dmg at all, on a note in the build workflow saying
hdiutil failed on the build machines. It does not; the run that note came from
had already failed for an unrelated reason, and the last error in the log was
read as the cause. The .dmg is back.
The app is ad-hoc signed but not notarized, so macOS says the developer cannot
be verified: right-click it and choose
Open, which offers the same dialog with a button that proceeds. Or:
xattr -dr com.apple.quarantine "/Applications/Zyvro Studio.app"Windows. The installer is not signed, so SmartScreen shows "Windows
protected your PC". Choose More info, then Run anyway.
Both warnings are accurate: neither platform can tell you this download came
from us rather than from someone who intercepted it. Every file's SHA-256 is at
the bottom of this page, and checking it is the only verification available
until the certificates are in place.
shasum -a 256 ~/Downloads/Zyvro* # macOS
certutil -hashfile Zyvro*.exe SHA256 # WindowsWhat is in it
- Large number of fix
Known limits
- Image nodes need a Google AI Studio or Black Forest Labs key once the free
daily allowance is spent - Video nodes need one of those two keys from the start: there is no free
allowance for video, and both backends bill by the second - The app has no self-updater, and the engine now ships inside it rather than
updating separately: a new engine means a new app - Nothing in the store is reviewed. Packs are signed now, but a signature says
who published something, not that it is safe — read a pack's Lua before
installing it; the app puts that one click from the listing - Publish to GitHub is not in the Git panel, and neither is an interactive
rebase or merge:pullis fast-forward only rather than leaving a repository
half-way through something this panel cannot finish
Checksums
| File | Size | SHA-256 |
|---|---|---|
Zyvro Studio-0.1.0-alpha.7-arm64.dmg |
134.4 MB | 7c4812c590598e6abdc2b1564b31ded8ed215579d69754ce04e75aa27c8b1d0a |
Zyvro Studio-0.1.0-alpha.7.dmg |
141.2 MB | 9e6a0c66bbec206d18feeb587bdf366771e15bae4f8ed725b2c29835caaa1828 |
Zyvro Studio v0.1.0-alpha.6
Zyvro Studio is a desktop workspace for Zyvro workflows: a file tree, an editor,
a terminal, source control, and an agent panel, with the graph editor opening in
a tab beside the code it acts on.
This is an alpha. It is built from three public repositories, it is not
signed, and the parts that touch a store are new. Read the two warnings below
before the first launch.
Why it exists
A ChatGPT or Claude subscription is not an API credential, and no hosted server
can use one. But the claude and codex command line tools on your own machine
are already signed in, so Zyvro Studio runs them as text providers. Your
subscription drives a workflow, and no credential ever reaches Zyvro.
Everything runs locally. Workflows live as files under .zyvro/ in the folder
you open, so they version alongside your code. The only traffic that leaves your
machine is the model call itself, and it goes straight to the provider.
First launch
macOS. Take the .dmg for your machine — arm64 for Apple Silicon, the
other for Intel — open it and drag the app to Applications. A .zip of the same
build is there too, for anyone who would rather not mount an image.
alpha.2 shipped without a .dmg at all, on a note in the build workflow saying
hdiutil failed on the build machines. It does not; the run that note came from
had already failed for an unrelated reason, and the last error in the log was
read as the cause. The .dmg is back.
The app is ad-hoc signed but not notarized, so macOS says the developer cannot
be verified: right-click it and choose
Open, which offers the same dialog with a button that proceeds. Or:
xattr -dr com.apple.quarantine "/Applications/Zyvro Studio.app"Windows. The installer is not signed, so SmartScreen shows "Windows
protected your PC". Choose More info, then Run anyway.
Both warnings are accurate: neither platform can tell you this download came
from us rather than from someone who intercepted it. Every file's SHA-256 is at
the bottom of this page, and checking it is the only verification available
until the certificates are in place.
shasum -a 256 ~/Downloads/Zyvro* # macOS
certutil -hashfile Zyvro*.exe SHA256 # WindowsNew in alpha.6
A browser, inside the IDE, that the agent drives. A page with its own
session, in a panel beside the code — so an agent checking its work never
touches the browser where you are signed in to everything. It opens a URL,
reads what is on the page, clicks, types, waits for text to appear, scrolls,
runs an expression, reads the console and the failed requests, and takes a
screenshot. Several pages at once, each named, each with its favicon in the
sidebar.
Right-click and Inspect, and Chromium's own developer tools open under the
page, on the element you pointed at: Elements with the live DOM, Console,
Network, Application with its cookies and its storage. Not an imitation and not
a second window to lose behind the others — the app is Chromium, so these are
its own tools.
The agent can write, and you say how much. Four levels, next to the
message box: read-only, ask before each change, the whole workspace, or
everything. Ask shows a card naming the tool and what it would do, with
Allow and Deny — the agent waits for you instead of stalling. The choice is
remembered per project, so a folder you set to one level opens at that level
next time, and the default is the whole workspace: an agent that cannot write
in the project you opened for it is an agent you have to argue with.
Any agent you start in the terminal gets this project's tools. Type
claude or codex in the shell and it comes up already knowing this project's
workflows, because the shell hands it the MCP servers on the way in. The port
and the token change at every start; nobody should have to go looking for them.
Search and replace, across the project or inside one file. ⇧⌘F searches
every file in the folder — plain text, whole word, or a regular expression with
$1 in the replacement — narrowed by the usual globs. Replace one occurrence,
all of them in one file, or all of them everywhere; the button says how many and
in how many files before you touch it. ⌘F stays in the file you are reading. What
gets replaced is what you were shown: each passage is checked against the file
again, and one that moved since the search is skipped rather than overwritten.
Arrow-up in the search box brings back what you searched before.
Photograph a panel, and hand it to someone. Click the camera, then the panel
you mean: the app knows where its own panels are, so the image stops at their
edges instead of a rectangle you dragged by hand, and the selection frame is not
in the picture. Then it asks what to do with it — keep it here, or upload it to
your Zyvro account and copy a public link.
Suggestions in the editor, off until you turn them on. Grey text at the
cursor, from a model of your choosing, and completing code is treated as its own
job rather than a flavour of text generation: it has its own list of providers,
because a chat model answers a sentence where the editor wants three characters.
A model on your own machine is usually the right one.
Video nodes. Text to video, on Veo through the Gemini API or on FLUX 3 Video
at Black Forest Labs, with an image on the input to open the shot. Resolution and
duration are settings on the node rather than a hidden default, because video is
the one thing here billed by the second and the price per second climbs with the
resolution. What one backend cannot do it refuses by name — Veo has no QHD step —
rather than quietly rendering something smaller than you asked for at the price
you asked for.
Runtime inputs are named where you can see them. The editor says which nodes
a run can fill by name, so a workflow you drive from the API or from an MCP tool
tells you what it expects instead of failing on a key nobody wrote down.
Fixed in alpha.6
The developer tools opened empty. Elements, Console, Network — the right
panel, all eight tabs, and nothing in any of them. Electron does not bridge its
inspector to a <webview>, so the tools were connected to nothing at all. They
are drawn in a native view now, which is the documented path and the one that
works.
A screenshot photographed its own overlay. The selection frame and "Click a
panel · Esc to cancel" were in the image: the overlay was dismissed and the
capture taken in the same tick, before the screen had been redrawn.
And then it left a hole where the devtools were. capturePage renders a
window's HTML, and a native view is drawn on top of it rather than inside — so
photographing the browser panel with the tools open produced a picture that
looked normal and was empty exactly where you wanted to look. The capture now
composes the native views back in.
Errors arrived with their plumbing. "Error invoking remote method
'shots:share': Error: Sign in to publish to the store." The sentence was written
to be read; the wrapper in front of it named a channel nobody needs to know, and
made a clear instruction look like a crash.
A cached run could replay the wrong result. A workflow driven by node id
rather than by input name folded the old value into its fingerprint, so a second
run with a different input was answered from the cache.
A Black Forest Labs key you had pasted never reached a run. The panel said
"Connected" — it reads what is stored — and the run said no key was configured,
because the step that folds a stored key into a run's configuration did not name
this one. Both statements were true and nothing connected them. Image nodes on
your own FLUX key were affected too; only the platform's key, which arrives by
another road, ever worked.
Smaller things. The shell greeted you in French in an English app. The agent's
message box overflowed its panel and left no room to type. Store thumbnails were
drawn by a copy of the web app's component instead of the component.
New in alpha.5
Models that never leave your machine. Ollama on this computer, LM Studio,
and any other server speaking the OpenAI chat API are providers now, for text
generation and for vision. They are configured by an address rather than by a
key, so the panel asks for one and fetches the model list from the server
itself — what is installed is your business and changes whenever you pull
something new. Asking a question about an image is the call where sending it
away is the whole cost, and a model on your own machine is the only way not to.
The same for images: any server speaking the OpenAI images API can back an
image node, generation and editing both. Editing used to be Gemini's alone on
the reasoning that it had no equivalent elsewhere. It has.
Providers are grouped by what they do. Image generation, vision and text
generation are three headings rather than two, because a provider that draws an
image cannot always read one — and when you have more than one for a job, you
say which is tried first.
The agent panel. Several agents at once, each in its own tab with its own
session; the session survives closing the app, so a conversation picks up where
it stopped. A model picker that shows the default it would have used. Images
you attach are handed to the CLI, which can actually open them. And what the
agent is doing is shown as it does it — the file it is reading, the command it
is running — rather than a spinner until it finishes.
Workflows move between projects. Import from another project on this
machine, share one by private link, and find everything you have put online
under My workflows.
Images in the chat. Paste, drop or pick one, and ask about it.
Fixed in alpha.5
Every image a run produced showed as broken. The app loads media from the
engine it started, on a port it only learns at launch, and the address being
built still carried the placeholder port — so generated ima...
Zyvro Studio v0.1.0-alpha.5
Zyvro Studio is a desktop workspace for Zyvro workflows: a file tree, an editor,
a terminal, source control, and an agent panel, with the graph editor opening in
a tab beside the code it acts on.
This is an alpha. It is built from three public repositories, it is not
signed, and the parts that touch a store are new. Read the two warnings below
before the first launch.
Why it exists
A ChatGPT or Claude subscription is not an API credential, and no hosted server
can use one. But the claude and codex command line tools on your own machine
are already signed in, so Zyvro Studio runs them as text providers. Your
subscription drives a workflow, and no credential ever reaches Zyvro.
Everything runs locally. Workflows live as files under .zyvro/ in the folder
you open, so they version alongside your code. The only traffic that leaves your
machine is the model call itself, and it goes straight to the provider.
First launch
macOS. Take the .dmg for your machine — arm64 for Apple Silicon, the
other for Intel — open it and drag the app to Applications. A .zip of the same
build is there too, for anyone who would rather not mount an image.
alpha.2 shipped without a .dmg at all, on a note in the build workflow saying
hdiutil failed on the build machines. It does not; the run that note came from
had already failed for an unrelated reason, and the last error in the log was
read as the cause. The .dmg is back.
The app is ad-hoc signed but not notarized, so macOS says the developer cannot
be verified: right-click it and choose
Open, which offers the same dialog with a button that proceeds. Or:
xattr -dr com.apple.quarantine "/Applications/Zyvro Studio.app"Windows. The installer is not signed, so SmartScreen shows "Windows
protected your PC". Choose More info, then Run anyway.
Both warnings are accurate: neither platform can tell you this download came
from us rather than from someone who intercepted it. Every file's SHA-256 is at
the bottom of this page, and checking it is the only verification available
until the certificates are in place.
shasum -a 256 ~/Downloads/Zyvro* # macOS
certutil -hashfile Zyvro*.exe SHA256 # WindowsNew in alpha.5
Models that never leave your machine. Ollama on this computer, LM Studio,
and any other server speaking the OpenAI chat API are providers now, for text
generation and for vision. They are configured by an address rather than by a
key, so the panel asks for one and fetches the model list from the server
itself — what is installed is your business and changes whenever you pull
something new. Asking a question about an image is the call where sending it
away is the whole cost, and a model on your own machine is the only way not to.
The same for images: any server speaking the OpenAI images API can back an
image node, generation and editing both. Editing used to be Gemini's alone on
the reasoning that it had no equivalent elsewhere. It has.
Providers are grouped by what they do. Image generation, vision and text
generation are three headings rather than two, because a provider that draws an
image cannot always read one — and when you have more than one for a job, you
say which is tried first.
The agent panel. Several agents at once, each in its own tab with its own
session; the session survives closing the app, so a conversation picks up where
it stopped. A model picker that shows the default it would have used. Images
you attach are handed to the CLI, which can actually open them. And what the
agent is doing is shown as it does it — the file it is reading, the command it
is running — rather than a spinner until it finishes.
Workflows move between projects. Import from another project on this
machine, share one by private link, and find everything you have put online
under My workflows.
Images in the chat. Paste, drop or pick one, and ask about it.
Fixed in alpha.5
Every image a run produced showed as broken. The app loads media from the
engine it started, on a port it only learns at launch, and the address being
built still carried the placeholder port — so generated images, workflow
thumbnails and chat attachments all pointed at nothing. A resolver for exactly
this existed and had no callers: every component used the shared one, which is
the point of sharing them. There is one function now, and a check in the build
that asserts the address it produces is the live port.
Runs were sent to a backend you never configured. Each router ended with a
fixed name, so somebody running LM Studio and holding no key at all was told to
go and get a key for a service they had never asked for. A rule that names
something usable is still obeyed; one that names something unusable now gives
way to what you actually have. Two things it was hiding: an installed claude
CLI was quietly elected as the default text backend — spending your
subscription because a binary exists is your decision, not ours — and every
install claimed to have an Ollama, because the default address was being counted
as a credential.
Fixed in alpha.4
The agent could not find claude or codex. Installed from the .dmg and
launched from the Finder, the app reported the CLI you use every day as "not on
your PATH" — true, and useless: it was on yours, and the app could not see it.
macOS launches an app from the Finder through launchd, which reads no shell
profile at all: no .zshrc, no .zprofile. The process starts with the system
default PATH, and everything a package manager or version manager installed is
missing from it. These two CLIs commonly live in ~/.local/bin, which is
nowhere in that list.
The app now asks your login shell what its PATH is and adopts it, and when a
tool is still missing it asks npm and brew where they put things rather than
guessing. On Windows it also looks for the extension, because the thing called
claude there is claude.cmd, which cannot be started without a command
interpreter.
This survived every test for one reason worth admitting: the app was always
launched from a terminal during development, where it inherited the right PATH.
The check that covers it now reproduces the Finder's environment instead.
New since alpha.1
Source control. A Git panel laid out the way editors lay it out: the message
box above the button, staged and unstaged in their own groups, the status letter
in the right-hand column, and the actions on the row you are pointing at. It
talks to the git on your machine rather than to a reimplementation, so what it
shows and what a terminal in the same folder shows cannot disagree. Stage,
unstage, discard, commit, branch, stash, tag, remotes, pull, push, and a
side-by-side diff on any changed file. Show Git Output keeps every command and
what git said back, which is the honest answer when something fails for a reason
no panel could have phrased.
Generate a commit message. A button in the message box reads the staged diff
and writes a subject line, using the claude or codex already installed on
your machine — one-shot, no tools, no MCP. Nothing goes to a server of ours and
there is no key to add. If neither CLI is installed, the button is not there.
Two modes. Dev is an editor: tabs, a shell underneath, the agent beside
you. AI is the agent: it takes the middle, there is no shell, and nothing is
open. Opening a file is what says you want to read it yourself, so that is when
the agent steps aside to the right.
Signed packs. A publisher's key is generated on their own machine and stored
only sealed under their password, which never reaches the server. Installing
verifies the signature, writes the key down the first time, and refuses the day
it changes — the same bargain SSH makes with host keys, and for the same reason.
A free daily allowance on the hosted store: ten image runs and ten model
calls a day per account, so a workflow you copy runs before you have added any
credential of your own.
What is in it
- The graph editor from the web app, hosted in an editor tab
- Read File and Write File nodes that act on the open project
- An agent panel running your own CLI, with this project's Zyvro tools over MCP
- Git, with a diff view, in a panel beside the file tree
- Node packs written in Lua, sandboxed: no filesystem, no network, no processes,
with a deadline, a memory ceiling and a model-call budget - A store for nodes and workflows, where installing a workflow pulls the nodes
it needs
Known limits
- Image nodes need a Google AI Studio or Black Forest Labs key once the free
daily allowance is spent - The app has no self-updater, and the engine now ships inside it rather than
updating separately: a new engine means a new app - Nothing in the store is reviewed. Packs are signed now, but a signature says
who published something, not that it is safe — read a pack's Lua before
installing it; the app puts that one click from the listing - Publish to GitHub is not in the Git panel, and neither is an interactive
rebase or merge:pullis fast-forward only rather than leaving a repository
half-way through something this panel cannot finish
Checksums
| File | SHA-256 |
|---|---|
Zyvro Studio Setup 0.1.0-alpha.5.exe |
bec2653def79324040adfd4105cdca31d4e0ad3df5375a8d91c43922d2981b56 |
Zyvro Studio-0.1.0-alpha.5-arm64-mac.zip |
a5a75a4ef99905ca756815a31558b233f7f34786ad67e25f5146add344e52bfb |
Zyvro Studio-0.1.0-alpha.5-arm64.dmg |
d7641936b904e1d60f45f3eb60e603d9e3fa730ff6b5058cc8212580843a4a37 |
Zyvro Studio-0.1.0-alpha.5-mac.zip |
6cdf988788d99b3f9f166f516ca9eb91d0fcd127a9843094323f944c2be73620 |
Zyvro Studio-0.1.0-alpha.5.dmg |
4157dcf63cb76b04866fcd232db59255d3c5838e63cad02c695444ccfeb8a3ba |
Zyvro Studio v0.1.0-alpha.4
Zyvro Studio is a desktop workspace for Zyvro workflows: a file tree, an editor,
a terminal, source control, and an agent panel, with the graph editor opening in
a tab beside the code it acts on.
This is an alpha. It is built from three public repositories, it is not
signed, and the parts that touch a store are new. Read the two warnings below
before the first launch.
Why it exists
A ChatGPT or Claude subscription is not an API credential, and no hosted server
can use one. But the claude and codex command line tools on your own machine
are already signed in, so Zyvro Studio runs them as text providers. Your
subscription drives a workflow, and no credential ever reaches Zyvro.
Everything runs locally. Workflows live as files under .zyvro/ in the folder
you open, so they version alongside your code. The only traffic that leaves your
machine is the model call itself, and it goes straight to the provider.
First launch
macOS. Take the .dmg for your machine — arm64 for Apple Silicon, the
other for Intel — open it and drag the app to Applications. A .zip of the same
build is there too, for anyone who would rather not mount an image.
alpha.2 shipped without a .dmg at all, on a note in the build workflow saying
hdiutil failed on the build machines. It does not; the run that note came from
had already failed for an unrelated reason, and the last error in the log was
read as the cause. The .dmg is back.
The app is ad-hoc signed but not notarized, so macOS says the developer cannot
be verified: right-click it and choose
Open, which offers the same dialog with a button that proceeds. Or:
xattr -dr com.apple.quarantine "/Applications/Zyvro Studio.app"Windows. The installer is not signed, so SmartScreen shows "Windows
protected your PC". Choose More info, then Run anyway.
Both warnings are accurate: neither platform can tell you this download came
from us rather than from someone who intercepted it. Every file's SHA-256 is at
the bottom of this page, and checking it is the only verification available
until the certificates are in place.
shasum -a 256 ~/Downloads/Zyvro* # macOS
certutil -hashfile Zyvro*.exe SHA256 # WindowsFixed in alpha.4
The agent could not find claude or codex. Installed from the .dmg and
launched from the Finder, the app reported the CLI you use every day as "not on
your PATH" — true, and useless: it was on yours, and the app could not see it.
macOS launches an app from the Finder through launchd, which reads no shell
profile at all: no .zshrc, no .zprofile. The process starts with the system
default PATH, and everything a package manager or version manager installed is
missing from it. These two CLIs commonly live in ~/.local/bin, which is
nowhere in that list.
The app now asks your login shell what its PATH is and adopts it, and when a
tool is still missing it asks npm and brew where they put things rather than
guessing. On Windows it also looks for the extension, because the thing called
claude there is claude.cmd, which cannot be started without a command
interpreter.
This survived every test for one reason worth admitting: the app was always
launched from a terminal during development, where it inherited the right PATH.
The check that covers it now reproduces the Finder's environment instead.
New since alpha.1
Source control. A Git panel laid out the way editors lay it out: the message
box above the button, staged and unstaged in their own groups, the status letter
in the right-hand column, and the actions on the row you are pointing at. It
talks to the git on your machine rather than to a reimplementation, so what it
shows and what a terminal in the same folder shows cannot disagree. Stage,
unstage, discard, commit, branch, stash, tag, remotes, pull, push, and a
side-by-side diff on any changed file. Show Git Output keeps every command and
what git said back, which is the honest answer when something fails for a reason
no panel could have phrased.
Generate a commit message. A button in the message box reads the staged diff
and writes a subject line, using the claude or codex already installed on
your machine — one-shot, no tools, no MCP. Nothing goes to a server of ours and
there is no key to add. If neither CLI is installed, the button is not there.
Two modes. Dev is an editor: tabs, a shell underneath, the agent beside
you. AI is the agent: it takes the middle, there is no shell, and nothing is
open. Opening a file is what says you want to read it yourself, so that is when
the agent steps aside to the right.
Signed packs. A publisher's key is generated on their own machine and stored
only sealed under their password, which never reaches the server. Installing
verifies the signature, writes the key down the first time, and refuses the day
it changes — the same bargain SSH makes with host keys, and for the same reason.
A free daily allowance on the hosted store: ten image runs and ten model
calls a day per account, so a workflow you copy runs before you have added any
credential of your own.
What is in it
- The graph editor from the web app, hosted in an editor tab
- Read File and Write File nodes that act on the open project
- An agent panel running your own CLI, with this project's Zyvro tools over MCP
- Git, with a diff view, in a panel beside the file tree
- Node packs written in Lua, sandboxed: no filesystem, no network, no processes,
with a deadline, a memory ceiling and a model-call budget - A store for nodes and workflows, where installing a workflow pulls the nodes
it needs
Known limits
- Image nodes need a Google AI Studio or Black Forest Labs key once the free
daily allowance is spent - The app has no self-updater, and the engine now ships inside it rather than
updating separately: a new engine means a new app - Nothing in the store is reviewed. Packs are signed now, but a signature says
who published something, not that it is safe — read a pack's Lua before
installing it; the app puts that one click from the listing - Publish to GitHub is not in the Git panel, and neither is an interactive
rebase or merge:pullis fast-forward only rather than leaving a repository
half-way through something this panel cannot finish
Checksums
| File | SHA-256 |
|---|---|
Zyvro Studio Setup 0.1.0-alpha.4.exe |
803421eb95b86c0e38b03b4e94d1ca212ca4f807ac7a43ec78c7db5a87e0299c |
Zyvro Studio-0.1.0-alpha.4-arm64-mac.zip |
4cf76362daf09eef50d875ae1e575382ad11433fbac76c4ab096abff1ebc1f60 |
Zyvro Studio-0.1.0-alpha.4-arm64.dmg |
89de864feec654a37702da3e55fcf82e2c8a986c6729615a9583de69d376a7d4 |
Zyvro Studio-0.1.0-alpha.4-mac.zip |
59fb629739132008b9bdad931a8423c785c72128f9455e0556000ade15d953d9 |
Zyvro Studio-0.1.0-alpha.4.dmg |
93a0332b0c6dc75a79aaa8698d5d688355391955b3232d2e3f2ae7f15d543938 |
Zyvro Studio v0.1.0-alpha.3
Zyvro Studio is a desktop workspace for Zyvro workflows: a file tree, an editor,
a terminal, source control, and an agent panel, with the graph editor opening in
a tab beside the code it acts on.
This is an alpha. It is built from three public repositories, it is not
signed, and the parts that touch a store are new. Read the two warnings below
before the first launch.
Why it exists
A ChatGPT or Claude subscription is not an API credential, and no hosted server
can use one. But the claude and codex command line tools on your own machine
are already signed in, so Zyvro Studio runs them as text providers. Your
subscription drives a workflow, and no credential ever reaches Zyvro.
Everything runs locally. Workflows live as files under .zyvro/ in the folder
you open, so they version alongside your code. The only traffic that leaves your
machine is the model call itself, and it goes straight to the provider.
First launch
macOS. Take the .dmg for your machine — arm64 for Apple Silicon, the
other for Intel — open it and drag the app to Applications. A .zip of the same
build is there too, for anyone who would rather not mount an image.
alpha.2 shipped without a .dmg at all, on a note in the build workflow saying
hdiutil failed on the build machines. It does not; the run that note came from
had already failed for an unrelated reason, and the last error in the log was
read as the cause. The .dmg is back.
The app is ad-hoc signed but not notarized, so macOS says the developer cannot
be verified: right-click it and choose
Open, which offers the same dialog with a button that proceeds. Or:
xattr -dr com.apple.quarantine "/Applications/Zyvro Studio.app"Windows. The installer is not signed, so SmartScreen shows "Windows
protected your PC". Choose More info, then Run anyway.
Both warnings are accurate: neither platform can tell you this download came
from us rather than from someone who intercepted it. Every file's SHA-256 is at
the bottom of this page, and checking it is the only verification available
until the certificates are in place.
shasum -a 256 ~/Downloads/Zyvro* # macOS
certutil -hashfile Zyvro*.exe SHA256 # WindowsNew since alpha.1
Source control. A Git panel laid out the way editors lay it out: the message
box above the button, staged and unstaged in their own groups, the status letter
in the right-hand column, and the actions on the row you are pointing at. It
talks to the git on your machine rather than to a reimplementation, so what it
shows and what a terminal in the same folder shows cannot disagree. Stage,
unstage, discard, commit, branch, stash, tag, remotes, pull, push, and a
side-by-side diff on any changed file. Show Git Output keeps every command and
what git said back, which is the honest answer when something fails for a reason
no panel could have phrased.
Generate a commit message. A button in the message box reads the staged diff
and writes a subject line, using the claude or codex already installed on
your machine — one-shot, no tools, no MCP. Nothing goes to a server of ours and
there is no key to add. If neither CLI is installed, the button is not there.
Two modes. Dev is an editor: tabs, a shell underneath, the agent beside
you. AI is the agent: it takes the middle, there is no shell, and nothing is
open. Opening a file is what says you want to read it yourself, so that is when
the agent steps aside to the right.
Signed packs. A publisher's key is generated on their own machine and stored
only sealed under their password, which never reaches the server. Installing
verifies the signature, writes the key down the first time, and refuses the day
it changes — the same bargain SSH makes with host keys, and for the same reason.
A free daily allowance on the hosted store: ten image runs and ten model
calls a day per account, so a workflow you copy runs before you have added any
credential of your own.
What is in it
- The graph editor from the web app, hosted in an editor tab
- Read File and Write File nodes that act on the open project
- An agent panel running your own CLI, with this project's Zyvro tools over MCP
- Git, with a diff view, in a panel beside the file tree
- Node packs written in Lua, sandboxed: no filesystem, no network, no processes,
with a deadline, a memory ceiling and a model-call budget - A store for nodes and workflows, where installing a workflow pulls the nodes
it needs
Known limits
- Image nodes need a Google AI Studio or Black Forest Labs key once the free
daily allowance is spent - The app has no self-updater, and the engine now ships inside it rather than
updating separately: a new engine means a new app - Nothing in the store is reviewed. Packs are signed now, but a signature says
who published something, not that it is safe — read a pack's Lua before
installing it; the app puts that one click from the listing - Publish to GitHub is not in the Git panel, and neither is an interactive
rebase or merge:pullis fast-forward only rather than leaving a repository
half-way through something this panel cannot finish
Checksums
| File | SHA-256 |
|---|---|
Zyvro Studio Setup 0.1.0-alpha.3.exe |
4392c4119ad4511a4752d8a9abc94672f6dce7966d0151c310300bc3cf2dbf47 |
Zyvro Studio-0.1.0-alpha.3-arm64-mac.zip |
087f84cce8fecf1e87f5cb11aa6e2a69a91040c558a2cfc98308ea84292a7053 |
Zyvro Studio-0.1.0-alpha.3-arm64.dmg |
9698378ea4e9224890579b260c44b9d5e1a8e30a0e95322dbdfaa6644ed655a9 |
Zyvro Studio-0.1.0-alpha.3-mac.zip |
c8e3fd278afeda73ebf5bb3a1e0161b91cf9f3e12199f04f8f65422b5dd08d56 |
Zyvro Studio-0.1.0-alpha.3.dmg |
78eb973831d637b6cb5837b52eee92728995b554b1b4562d18bdb527df8d6679 |
Zyvro Studio v0.1.0-alpha.2
Zyvro Studio is a desktop workspace for Zyvro workflows: a file tree, an editor,
a terminal, source control, and an agent panel, with the graph editor opening in
a tab beside the code it acts on.
This is an alpha. It is built from three public repositories, it is not
signed, and the parts that touch a store are new. Read the two warnings below
before the first launch.
Why it exists
A ChatGPT or Claude subscription is not an API credential, and no hosted server
can use one. But the claude and codex command line tools on your own machine
are already signed in, so Zyvro Studio runs them as text providers. Your
subscription drives a workflow, and no credential ever reaches Zyvro.
Everything runs locally. Workflows live as files under .zyvro/ in the folder
you open, so they version alongside your code. The only traffic that leaves your
machine is the model call itself, and it goes straight to the provider.
First launch
macOS. The download is a zip. Unzip it in Finder, or with ditto -x -k,
and move the app to Applications. The app is ad-hoc signed but not notarized,
so macOS says the developer cannot be verified: right-click it and choose
Open, which offers the same dialog with a button that proceeds. Or:
xattr -dr com.apple.quarantine "/Applications/Zyvro Studio.app"Windows. The installer is not signed, so SmartScreen shows "Windows
protected your PC". Choose More info, then Run anyway.
Both warnings are accurate: neither platform can tell you this download came
from us rather than from someone who intercepted it. Every file's SHA-256 is at
the bottom of this page, and checking it is the only verification available
until the certificates are in place.
shasum -a 256 ~/Downloads/Zyvro* # macOS
certutil -hashfile Zyvro*.exe SHA256 # WindowsNew since alpha.1
Source control. A Git panel laid out the way editors lay it out: the message
box above the button, staged and unstaged in their own groups, the status letter
in the right-hand column, and the actions on the row you are pointing at. It
talks to the git on your machine rather than to a reimplementation, so what it
shows and what a terminal in the same folder shows cannot disagree. Stage,
unstage, discard, commit, branch, stash, tag, remotes, pull, push, and a
side-by-side diff on any changed file. Show Git Output keeps every command and
what git said back, which is the honest answer when something fails for a reason
no panel could have phrased.
Generate a commit message. A button in the message box reads the staged diff
and writes a subject line, using the claude or codex already installed on
your machine — one-shot, no tools, no MCP. Nothing goes to a server of ours and
there is no key to add. If neither CLI is installed, the button is not there.
Two modes. Dev is an editor: tabs, a shell underneath, the agent beside
you. AI is the agent: it takes the middle, there is no shell, and nothing is
open. Opening a file is what says you want to read it yourself, so that is when
the agent steps aside to the right.
Signed packs. A publisher's key is generated on their own machine and stored
only sealed under their password, which never reaches the server. Installing
verifies the signature, writes the key down the first time, and refuses the day
it changes — the same bargain SSH makes with host keys, and for the same reason.
A free daily allowance on the hosted store: ten image runs and ten model
calls a day per account, so a workflow you copy runs before you have added any
credential of your own.
What is in it
- The graph editor from the web app, hosted in an editor tab
- Read File and Write File nodes that act on the open project
- An agent panel running your own CLI, with this project's Zyvro tools over MCP
- Git, with a diff view, in a panel beside the file tree
- Node packs written in Lua, sandboxed: no filesystem, no network, no processes,
with a deadline, a memory ceiling and a model-call budget - A store for nodes and workflows, where installing a workflow pulls the nodes
it needs
Known limits
- Image nodes need a Google AI Studio or Black Forest Labs key once the free
daily allowance is spent - The app has no self-updater, and the engine now ships inside it rather than
updating separately: a new engine means a new app - Nothing in the store is reviewed. Packs are signed now, but a signature says
who published something, not that it is safe — read a pack's Lua before
installing it; the app puts that one click from the listing - Publish to GitHub is not in the Git panel, and neither is an interactive
rebase or merge:pullis fast-forward only rather than leaving a repository
half-way through something this panel cannot finish
Checksums
| File | SHA-256 |
|---|---|
Zyvro Studio Setup 0.1.0-alpha.2.exe |
ca818138f4a6b89060f9cf83f45ed7aabe1675aaeeac8aafcb9a4c87ac357188 |
Zyvro Studio-0.1.0-alpha.2-arm64-mac.zip |
cb488370cedbe0d64bf91d8dfe1b00e76f395ddd361472ace1e177ebbf6b137c |
Zyvro Studio-0.1.0-alpha.2-mac.zip |
42b143da0ed98a35acf8bd00aceee56fe51043faa1f8538667a9010ffe030c43 |
Zyvro Studio v0.1.0-alpha.1
Zyvro Studio is a desktop workspace for Zyvro workflows: a file tree, an editor,
a terminal, and an agent panel, with the graph editor opening in a tab beside
the code it acts on.
This is an alpha. It is built from three public repositories, it is not
signed, and the parts that touch a store are new. Read the two warnings below
before the first launch.
Why it exists
A ChatGPT or Claude subscription is not an API credential, and no hosted server
can use one. But the claude and codex command line tools on your own machine
are already signed in, so Zyvro Studio runs them as text providers. Your
subscription drives a workflow, and no credential ever reaches Zyvro.
Everything runs locally. Workflows live as files under .zyvro/ in the folder
you open, so they version alongside your code. The only traffic that leaves your
machine is the model call itself, and it goes straight to the provider.
First launch
macOS. The download is a zip. Unzip it in Finder, or with ditto -x -k,
and move the app to Applications. The app is ad-hoc signed but not notarized,
so macOS says the developer cannot be verified: right-click it and choose
Open, which offers the same dialog with a button that proceeds. Or:
xattr -dr com.apple.quarantine "/Applications/Zyvro Studio.app"Windows. The installer is not signed, so SmartScreen shows "Windows
protected your PC". Choose More info, then Run anyway.
Both warnings are accurate: neither platform can tell you this download came
from us rather than from someone who intercepted it. Every file's SHA-256 is at
the bottom of this page, and checking it is the only verification available
until the certificates are in place.
shasum -a 256 ~/Downloads/Zyvro* # macOS
certutil -hashfile Zyvro*.exe SHA256 # WindowsWhat is in it
- The graph editor from the web app, hosted in an editor tab
- Read File and Write File nodes that act on the open project
- An agent panel running your own CLI, with this project's Zyvro tools over MCP
- Node packs written in Lua, sandboxed: no filesystem, no network, no processes,
with a deadline, a memory ceiling and a model-call budget - A store for nodes and workflows, where installing a workflow pulls the nodes
it needs
Known limits
- Image nodes need a Google AI Studio key; there is no CLI path for them
- The app has no self-updater. The engine updates itself, the app does not.
- Nothing in the store is reviewed, and packs are not signed. Read a pack's Lua
before installing it; the app puts that one click from the listing.
Checksums
| File | SHA-256 |
|---|---|
Zyvro.Studio-0.1.0-alpha.1-arm64-mac.zip |
878bf4a857e1b7946afd836d1830dbe5e521bef2869ff54797abcb87736e24d6 |
Zyvro.Studio-0.1.0-alpha.1-mac.zip |
a7a9eb0307c5320a1e8c7ff79daa3d73fba3bebb9013e0f3827b9f9014c71632 |
Zyvro.Studio.Setup.0.1.0-alpha.1.exe |
b5c5f9b9c5714ccfed58c652d590bcc0f0ac0324e6749c6632dd6c1c35462a65 |