ADE · Agent Development Environment
Zyvro — More than prompts

An IDE built around the agent, not beside it

Claude Code, Codex and Qwen run inside the window, on your machine — with your files, your terminal, your git and your own keys. Beside them, a canvas where you wire the workflows they can call.

The editor

It is an IDE.
It just knows what your workflows are.

Your files, a real shell, git, and the canvas in the same window. One switch changes who the window is for: write the code yourself, or ask for it. Nothing else moves — the same project, the same workflows, the same models.

Dev mode: the file tree and this project's workflows on the left, a file open in the editor, a shell underneath, the agent panel on the right

Editor, shell underneath, agent beside.

The canvas

Wire the models together.
Run it as many times as you like.

Text, image and vision models on one canvas, each node carrying what it produced. A workflow is a file in your project — it runs from here, from the chat, from an API call, or from an agent.

A workflow on the canvas: two prompts and two reference images feeding an image generator, then background removal, then a preview

Your own models

Nothing has to leave your machine.

Ollama on this computer, LM Studio, or any server speaking the OpenAI API — an address and a model, no key to paste. The list of models comes from the server itself, because what is installed is your business. It is also the only way to ask a question about an image without the image going anywhere.

Choosing a local model: an address, and the models the server reports

The agent

Your subscription, not another bill.

The panel runs the claude or codex CLI already signed in on your machine, in your project folder. A ChatGPT or Claude subscription drives a workflow with no API key — Zyvro never sees a token. It holds this project's tools, so it can list your workflows, read a graph, run one and read what came back.

The agent panel, running the CLI installed on this machine

Permissions

You decide what it may do.

Per conversation, next to what you are about to ask. Workspace, and it works in the project without asking. Ask, and every action arrives with its command or its path, to allow or deny. Read only, and it looks without touching. YOLO, if you mean it.

The four permission levels in the agent composer: Workspace, Ask, Read only, YOLO

Agent tools

Every shell hands over the project.

Open a terminal and it already knows where this project's engine listens and what opens it: the addresses, the tokens, a ready-made mcp.json for any client, and a zyvro-mcp command that starts claude or codex already wired in. The port and the token change at every start — nobody should have to go looking for them.

A shell in the IDE listing the project's MCP servers, their addresses and the config file

The browser

It has its own. Yours keeps its tabs.

A page inside the IDE, in a session of its own. An agent opens it, reads what is on it, clicks, types, waits for a text to appear and photographs it — without touching the browser where you are signed in to everything. Right-click and Chromium's own devtools open under the page: Elements, Console, Network, Application. It is Chromium, because the app is.

Two pages open in the IDE's own browser, listed in the sidebar, with the address bar and the shell below

Under the page

Chromium's own devtools, docked.

Right-click, Inspect, and the real front-end opens under the page — Elements on 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: it is the inspector the app already carries, drawn in the panel, next to the code that produced the page.

The IDE's browser showing a page, with Chromium's Elements panel and its DOM tree docked underneath

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 — and you narrow it with the usual globs. Replace one occurrence, or all of them in one press: the button says how many and in how many files before you touch it. ⌘F stays in the file you are reading, in the editor's own bar. 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.

The search panel: a query, a replacement, file filters, and results grouped by file with each match highlighted

Inline completion

Grey text, from a model you chose.

Ghost text in the editor, from a provider you picked for that job — and it is a list of its own, because the models that fill in the middle of a file are not the ones that chat. It waits for a pause before asking and drops the request the moment you type again, so a suggestion that arrived late never lands on a cursor that moved. Off until you turn it on: every pause spends an account.

The code completion providers in the desktop settings, separate from text generation

Source control

Git, beside the canvas.

Pull, stage, commit, push, branch, stash, tag, and a side-by-side diff on any changed file. The commit message writes itself when you want it to: the claude or codex CLI already signed in on this machine reads what is staged and proposes one line, which you edit before it goes anywhere — no key, no other bill. Conflicts get their own group at the top, because they are the only thing you cannot clear by clicking a plus. It drives the git on your machine rather than a reimplementation of it, so this panel and a terminal in the same folder cannot disagree.

The source control panel: a commit message written by the agent, and the changed files

Screenshots

Show someone what you are looking at.

Click the camera, then the panel you mean. The app knows where its own panels are, so the image stops on their edges instead of a rectangle you dragged by hand — and the selection frame disappears before the shot, so it is not in the picture. Then it asks what to do with it: Download keeps it on this machine, Share uploads it to your Zyvro account and copies a public link. That is the whole gesture behind “look at this”, whether it is the thing you just built or the error you cannot explain.

The whole gesture: the panels outline themselves, one is picked, the image appears in a dialog, and the link is copied

The store

Start from one that already ran.

Workflows other people published, each shown as what went in and what came out — because a name and a description do not tell you whether it does what you want. Installing pulls the node packs it needs. They run on your machine, in the sandbox, and you can read the source before installing anything.

Browse the store
The store: published workflows, each with its input and its result

Made with Zyvro

Published workflows, each shown as what went in and what came out.

Browse all

Every node

The whole palette, so you can see what a workflow can be made of.

Text InputInput
Static text or runtime input
Image InputInput
Uploaded image or runtime input
Read FileInput
Read a file from the project folder (desktop only)
Merge TextUtility
Concatenate upstream texts
LLMModel
Text completion via Ollama, Claude or OpenAI
Generate ImageModel
Text -> image, on Gemini or FLUX
Generate VideoModel
Text -> video, on Veo or FLUX
Edit ImageModel
Edit an image with a prompt
Remove BackgroundModel
AI two-pass matte, or programmatic color-range removal
Rotate ImageUtility
Rotate an image by 90-degree steps
Flip ImageUtility
Mirror an image horizontally or vertically
Vision / JudgeModel
Analyze image(s) with a VLM
BrainAgent
Agent that calls connected nodes as tools
Workflow ToolsAgent
Gives a Brain the MCP tools: list, inspect and run your other workflows
PreviewOutput
Render a result in the node
3D Voxel PreviewOutput
Cube or sphere textured with the connected images
Write FileOutput
Write the result into the project folder (desktop only)
OutputOutput
Mark a workflow result