For Mac · Beta
An AI assistant that runs on your own models, and asks before it acts.
Zikporx AI is a Mac app for chatting with AI and working on code. It talks to the AI models you run yourself with Ollama, so your conversations and your code stay on your computer. When it wants to run a command, change a file or use your browser, it shows you exactly what first.
macOS · Apple Silicon and Intel · Needs Ollama and a model · Signed and notarized by Apple
In one minute
What you get
Chat
Ask questions, attach files and images, dictate by voice, and keep projects and server connections together. Each answer shows how long it took and how many tokens it used.
Code
An editor with an AI that proposes changes as a before-and-after you accept or reject. Every change is checked before you see it, and your project's own build and tests can be run to find real errors.
Tools
Let the assistant use your browser, files and servers through a tool system (MCP). Anything that changes something shows an approval card with the exact action, and every call is written to an activity log.
Private by design
No account and no cloud model. The AI runs on your Mac or on a machine you control. Your chats are stored on your computer.
You stay in control
Nothing that changes your files, runs a command or controls your browser happens without your approval. Sending an email is switched off unless you turn it on.
Honest about limits
It is only as capable as the model you run. Small models are fine for chat and single steps and weaker at long chains of actions. This page says where.
How it works
Three pieces: the app, a model server, and tools
Zikporx AI is the part you see. It does not contain an AI model. It sends your message to a model server you run, and shows you what comes back.
1. Zikporx AI (this Mac)
The chat and code windows, your projects, the approval cards and the activity log. It builds each request and decides what is allowed to run.
2. Your model server
Ollama, on this Mac or on another machine on your network. It runs the model (for example a Qwen, Llama or Mistral model) and streams the answer back.
3. Tools (MCP, optional)
Add-ons you switch on: your Safari or Chrome, the built-in background browser for searching the internet, files in a folder, or any MCP tool you add. Each one asks first.
What happens when you send a message
- You type something.The app attaches what the model needs: the earlier messages of this chat (shortened when they get long), any files you attached, your project's instructions if a project is open, and a short description of the tools that are switched on.
- The request goes to your model server.Over your own network or this Mac. Nothing goes to a Zikporx server, because there isn't one.
- The answer streams back.You see it appear word by word, with a running timer and token count. When it ends, the app shows the exact time and the number of tokens the model reports.
- The app reads the answer for actions.If the model asked to use a tool, run a command, or change a file, the app turns that into an action. Plain text is just shown.
- Actions that change something wait for you.You get an approval card with the exact command, file text or tool arguments. You can approve, deny, or type what to change instead.
- The result goes back to the model.It reads what happened and either continues or tells you the outcome. Everything that ran is written to the activity log.
Built into the app itself, no MCP and no add-ons needed: chat with attachments and voice, projects, the code editor with checked edits, connecting to your servers over SSH and running commands on them (including installing software and deploying), the terminal, git and deploys, and the activity log. MCP tools are an optional extra on top of these.
Why this matters: the model never touches your Mac directly. It can only ask. The app is the one that decides, checks and runs, and it asks you. That is the whole safety model.
A tool request, step by step
{ "browser": "safari", "url": "https://apple.com" }Short, clear requests like this one are recognised by the app itself, so they work even with a small model. Longer jobs go to the model, which then has to choose the right tools in the right order. That second part is where model size matters (see What to expect).
Chat
Everyday AI chat, with your projects and servers
Conversations
- Pick any model your Ollama server has, and download or delete models from the model menu.
- Attach images, text and code files, and PDFs. The text of a file goes to the model; it is not shown in your message bubble.
- Dictate with the microphone. Speech is turned into text by a speech model that ships inside the app, on your Mac.
- Each reply ends with its time and token count, so you can see what a long chat costs your model.
- Dark and light themes, a resizable sidebar, and a Help & FAQ page.
Projects
- Open a folder as a project. The assistant sees the file list and can read files you point it at.
- Instructions: write rules once ("use tabs", "never touch /config") and they are included in every request for that project.
- Git: see changes and commit with an AI-suggested message. Committing never pushes.
- Deploys: a history of what was deployed, with rollback.
- Terminal: run commands in the project folder, with live output.
Servers over SSH
- Built in. This works on its own: no MCP server or other add-on is involved.
- Connect to a server by chatting ("connect to my VPS") or from the Servers list.
- Ask for a job in plain words, such as "install nginx and serve this folder". It plans the commands, shows each one, and runs them over SSH only after you approve. It also has step-by-step recipes for common installs and deploys.
- Passwords are encrypted with your Mac's keychain, not stored as text.
- Commands the assistant wants to run on the server appear on an approval card. Risky ones, such as
rm -rf, are flagged in red. - Package-manager prompts like "Continue? [Y/n]" are answered automatically so a command can't hang forever.
Optional extras
- Built-in background browser: the assistant can search the internet and read pages in a hidden browser inside the app, with no sign-ins. See Tools.
- Web search (off by default): the assistant can look something up and read a page. You can point it at your own SearXNG server.
- Offline demo mode: built-in sample answers, no server needed, to look around the app.
- Activity log: every command, edit, tool call and server connection, with filters and export.
Code Beta
An editor where every AI change is checked before you see it
Open a project folder, open a file, and ask for what you want in plain words. The AI sees the file (with line numbers, so "line 62" means line 62) and proposes edits.
How a change reaches your file
- You ask."In
trade(), return early when the balance is zero." You can also select code first to send only that part. - The AI answers with edits, each pointing at exact lines of the exact file.
- The app checks every edit before showing it.The result must still parse. For Python and JavaScript it also checks for undefined names and code that can never run, and warns when a change removes behaviour or breaks other files that depend on it.
- You see a red-and-green before/after and choose Accept or Reject, one change at a time or all at once.
- A safety copy is kept.Before the first AI edit of a file, a checkpoint is saved in the project's own hidden git history, and an Undo button restores the previous version.
A change that would break the file is refused. You see the reason and the line, and nothing is touched. A change that looks risky shows a warning above the diff.
Finding real problems
Review
Reads the whole file in parts. First the real checks run (syntax, undefined names, unreachable code, bad imports). Then the AI adds its own findings. Problems the checks prove come first; the AI's findings are labelled as opinions, and ones the code already seems to handle carry a note, so you can decide.
What's missing
Looks at the open file only. The top list is checked by code, not guessed: names that are never defined, empty functions, errors that are caught and ignored, divisions with no zero check, unfinished TODO notes. Below it, AI suggestions sit in a separate, clearly labelled list.
Check & Fix
Runs your project's own tools (build, type check, tests, linters) in a temporary copy, and lists the real errors with file and line. Click one to jump to it, or ask the AI to fix them as normal accept/reject edits, then run again to see how many are left.
Run & test change
▶ Check runs the open file in a temporary copy of your project. ▶ Test change runs the file before and after a waiting change, and tells you if the change fixes the failure or makes it fail.
Languages
Every proposed edit gets a syntax check in these languages: Python, JavaScript, TypeScript and TSX, Java, C#, C and C++, PHP, Go, Rust, Kotlin, Swift, SQL, shell, Ruby, Dart and CSS. The app uses the language's own tool when it is installed on your Mac (for example gofmt, php, ruby, bash, swiftc, clang or the project's TypeScript) and a built-in bracket and string scanner when it is not. Check & Fix also detects npm scripts, Python tests, Go, Cargo, Swift, Gradle and Maven, .NET, CMake and Make.
Also in the Code tab
- A terminal for the project, with live output and a Stop button.
- Search across the project, with highlights and case or regex options.
- A server mode: open a folder on a server over SSH and work on it the same way.
Tools
Add tools: your browser, your files and other programs
Tools use the open standard called MCP (Model Context Protocol). A tool is a small program that offers actions such as "read this folder" or "open this page". You choose which ones are on. They are an add-on: chat, projects, the code editor and server connections over SSH work without any of them.
Tools on your network
Add any MCP server by address, with a token that is stored encrypted. Good for files, git or databases on another machine.
Tools on this Mac
Programs the app starts for you (files in a folder, memory, the browser tool). They sit behind one master switch, Mac, which is off by default and shows a warning when you turn it on.
Built-in background browser
Search the internet and read pages without opening your own browser. A hidden window inside the app, with an empty profile, that works even when the Mac switch is off. Details below.
The built-in background browser: search the internet without your own browser
Zikporx AI has its own built-in background browser, a browser that is part of the app and works in the background. It is a hidden window inside the app, so you don't see it and it doesn't touch your Safari or Chrome. Ask "search the web for gaming laptops" or "open this page and tell me what it says", and the assistant uses it to look things up and read them.
What it can do
- Search the web and show you the top results.
- Open a page and read all of its text, scrolling down to load the whole page.
- List the form fields and buttons on a page, fill a field and click a button (each step asks first).
- Find, then show you: it can find the best result in the background and then open it in your own browser, or just give you the link.
- It works even when the Mac switch is off, because it is part of the app and starts no other program.
How it is kept safe
- Starts empty and forgets. It has its own temporary profile with no cookies and no sign-ins. Its data is cleared when it closes or sits idle for five minutes.
- Web addresses only. It opens http and https pages, and it will not open addresses on your own computer or your local network.
- No downloads, no pop-ups, no permission requests (camera, location and similar are refused).
- Never fills passwords, cards or IDs, and never clicks payment, purchase or Send buttons.
- Text on a page is treated as data. The assistant is told never to follow instructions found inside a page.
| Built-in background browser | Safari & Chrome tool | |
|---|---|---|
| Where it works | A hidden window inside the app | The tab you already have open |
| Signed in to your accounts? | No, always empty | Yes, it uses your own sign-ins |
| Good for | Searching, reading public pages, quick lookups | Your email, your logged-in sites, anything you need to see |
| Needs the Mac switch | No | Yes |
| Needs macOS permission | No | Yes: Automation, and "Allow JavaScript from Apple Events" for reading and filling |
| Can you watch it? | No, it is hidden | Yes, in your own Safari or Chrome |
The Browser tool for Safari and Chrome
This tool controls the tab you already have open, with your own sign-ins, using macOS Apple Events. It can:
| Action | What it does | Asks first? |
|---|---|---|
| Read | Reads the open page (it scrolls through the whole page, then goes back to where you were) and lists open tabs. | No: read-only |
| Open, search | Opens an address, or types a query into the page's search box and submits it in one step. | Yes |
| Fill, click | Lists the form fields and buttons by number, fills a field, clicks a button. | Yes |
| Find an email | Finds an email by sender or subject in your mail page and opens it, returning its text. | Yes |
| Write a reply | Puts a drafted reply at the top of the reply box, keeping the quoted message and your signature. It never presses Send. | Yes |
| Send an email | Presses Send, only if you switched this on in Settings, and only after a card showing the recipient, subject and text. It refuses if the open draft is not the one you approved. | Yes, and off by default |
What it will never do on its own: fill a password, card number or ID field; click a payment, purchase, transfer or delete-account button; or press Send unless you enabled sending and approved the card.
Setup is one time: macOS asks whether Zikporx AI may control Safari or Chrome, and to read pages or fill forms Safari needs Develop → Allow JavaScript from Apple Events. The app's tool screen has a checklist and a Check button for this.
Example: reply to an email
- "Find the email from John and open it."The app looks in the visible mail list, or uses the mail search box, then opens the message and reads its text.
- "Write a reply."The model drafts from the text it just read, addressed to the sender, with no placeholders.
- "Put it in the reply box."The draft is typed into the open reply. You read it in your mail page.
- You press Send, or, if you turned sending on, you approve a card that shows who it goes to and what it says.
Safety and privacy
What it can see, what it can do, and who sees what
Where your data goes
- Your messages go to the Ollama server you set, nowhere else.
- Chats, projects and settings are stored on your Mac.
- There is no Zikporx account and no cloud AI, and the app collects no analytics or usage data.
- Web search and the built-in background browser's search use DuckDuckGo. Remote tools you add receive only what you send them. These are the only requests the app makes beyond your model server and the pages you ask it to open.
What needs your approval
- Any command, on this Mac or a server.
- Any file the assistant wants to write.
- Any tool call that is not read-only.
- Each card shows the exact text. You can approve, deny, or write what to change.
Guard rails inside the app
- Text found on a web page or in an email is treated as data. The model is told never to follow instructions that appear inside it.
- Passwords, card numbers and ID fields can't be filled by the tools.
- Test runs happen in a temporary copy of your project, without your
.envfiles and with a separate home folder. - A few clearly dangerous commands (such as
sudoorgit push) are refused in test runs.
An honest note on limits
- Test runs are not a sandbox: a script you run can still use the internet and your account. Only run code you trust.
- The model can be wrong. Read what it proposes, especially commands and edits.
- Passwords for servers are encrypted with your keychain, but anyone who can use your logged-in Mac can use them.
Everything is logged. The Activity log lists each command, edit, tool call, deploy and connection, with the time, the result and a short excerpt of the output, so you can always see what happened.
What to expect
The model decides how capable it is
Zikporx AI is only as capable as the model behind it. This is the single most important thing to know before you try it.
| Model size | Good for | Struggles with | Memory (4-bit) |
|---|---|---|---|
| 7–14B | Chat, explaining code, single clear tool calls, short edits | Chains of several tool steps, keeping a goal over a long task | about 5–10 GB |
| 24–32B | Reliable multi-step tool use (about 3–6 steps), better reviews | Very long jobs | about 16–22 GB |
| 70B and up | Planning and recovering from errors | Needs a powerful machine | about 40–48 GB |
- Pick a model made for instructions and tools. A model tuned only for code is weaker at choosing tools than a general instruct model of the same size.
- Review shows the model's opinions. The checks it runs itself are exact; its other findings can be wrong, which is why they are labelled and can carry a "likely a false alarm" note.
- Web pages change. Gmail and similar pages are complex, and a page redesign can break a tool step. When that happens the tool says so; it does not guess.
- The Code tab is in beta. Keep a backup or use git for projects that matter. Changes are shown before they are applied, and an Undo is kept.
Download
Get the beta
Install in four steps
- Open the DMGand drag Zikporx AI to your Applications folder. It is signed and notarized by Apple, so it opens without a warning.
- Install Ollamafrom ollama.com on this Mac, or on another machine on your network.
- Download a model.For example, in Terminal:
You can also download models from the model menu inside the app.ollama pull qwen3:14b - Open Zikporx AIand set the server address in Settings (the default is
http://localhost:11434). When it says "Ollama connected", you are ready to chat.
Known issues in this beta
- Multi-step browser and email tasks can fail with small models. Short requests such as "open apple.com in Safari" work.
- The Code tab's deeper checks cover Python and JavaScript. Other languages get syntax checks and their own build tools.
- There is no automatic update yet. You download each new version yourself.
Feedback
Tell us what broke and what confused you in the feedback form. The Activity log has a copy option, so you can attach what happened.
Feedback
Tell us what you think
Found a bug, want a feature, or just want to say what worked? Write it here. It is a beta, and every note helps decide what gets fixed first.
What people are saying
Talk with other people
Community forum
Ask questions and share how you use Zikporx AI. Link to your forum or discussions page
Report a security problem
If you found something that could put someone's data at risk, please write privately first. Security contact email
Questions
FAQ
Does it send my code or chats to the cloud?
No. Messages go to the Ollama server you choose. The app has no account system and no cloud model. If you enable web search or add a remote tool, those specific requests go to the address you set.
Do I need to pay for a model or an API key?
No key is needed. You run open models yourself with Ollama. The cost is the computer that runs them.
Which Mac do I need?
Any Mac that runs the app (Apple Silicon or Intel). The model needs memory: a 14B model wants roughly 10 GB, a 32B model roughly 20 GB. It can run on a different machine on your network.
Can it send emails or buy things for me?
The browser tool refuses payment, purchase, transfer and delete-account buttons, so you complete those yourself. It can write an email reply into your mail page and leave it for you to send. Sending is a separate setting that is off by default, and it requires you to approve a card that shows the recipient, subject and text each time.
Can it break my project?
It can only change files after you accept a change. Every change is checked first and a checkpoint plus an Undo is kept. Keep git or a backup for important work, because the Code tab is in beta.
Why did it say it can't open Safari?
With a small model, it sometimes answers from habit instead of using the tool. Say what you want in a short, clear sentence, for example "open apple.com in Safari", and check that the Mac switch is on and Safari is allowed in System Settings → Privacy & Security → Automation.
Is it an autonomous agent?
No. It acts step by step and asks before anything that changes something. That is deliberate. A longer, more autonomous mode needs a stronger model and more safeguards, and is not part of this beta.
What does it cost?
Add your pricing or “free during the beta” here.