WidelAI for Mac Is Here, With a Coding Agent That Runs Locally

WidelAI is now a Mac app. You can download it today, sign in with the account you already use on the web, and carry on the same conversations. It is a native app written in Swift and SwiftUI — not the website in a wrapper — and it does one thing the web version cannot: it works on your actual project.

The Mac app includes a coding agent. Point it at a folder and it reads the files, proposes edits, and runs commands in your own shell. Not in a container that cannot see your toolchain, and not on a copy of your repository uploaded somewhere. On the files that are on your disk, with the versions of things you actually have installed.

That is a meaningful amount of trust to ask for, so most of this post is about the parts that constrain it: what the agent can reach, how you review what it did, and how to undo it.

What You Get

• Chat with models from OpenAI, Anthropic, Google, Moonshot, and Zhipu from one picker
• Switch models inside a conversation and keep the full history
• A coding agent that reads and edits files in a project folder you choose
• Shell commands run locally, so build tools, linters, and test runners are the ones you have installed
• Every change as a reviewable diff, file by file, before you accept it
• Checkpointed runs — revert one edit or roll back an entire session
• Several projects at once, each with its own session and history
• Streaming replies with properly rendered markdown and code blocks
• File and image attachments on any message
• Your session in the macOS Keychain, not a plain file
• Signed and notarized by Apple, with updates verified before they install

Model availability depends on your plan, and an active plan is required to generate responses.

The Agent Works on Your Machine, Not a Copy of It

Most AI coding tools fall into one of two camps. Browser-based ones cannot see your code unless you paste it in, which means they are guessing about everything you did not paste. Container-based ones get a checkout, but not your environment — not your Node version, not your local database, not the environment variable that explains why the test fails on your machine and nowhere else.

A local agent avoids both problems. It reads the file, it runs the command, it sees the real error. When it says the tests pass, they passed on your machine.

The trade-off is obvious and worth stating plainly: something is editing your files and running commands as you. Here is what bounds that.

It is scoped to one folder

You open a project. The agent can read and write inside that folder and nothing else — path traversal out of the workspace is refused, not merely discouraged. Open a different project and you get a separate session with its own workspace.

Every change arrives as a diff

The agent does not silently rewrite files and tell you it is done. Each edit is presented as a diff you read before accepting. This is the step worth not skipping: reviewing a twenty-line diff takes fifteen seconds and is the difference between using an agent and hoping.

Every run can be reverted

Runs are checkpointed. If the agent went down the wrong path four files in, roll the session back. Undo a single edit, or all of them. There is no state where your only option is to reach for git checkout and hope you committed recently.

The honest summary: the agent is fast at the tedious parts and occasionally confidently wrong about the important ones. The diff review and the revert button are what make that an acceptable trade rather than a gamble.

Switch Models Mid-Conversation

The feature that changes how people work is the same one as on Android: you can change model in the middle of a conversation and the thread comes with you.

1. Ask a question of the model you reach for by default.
2. Switch models. Ask the same thing again.
3. Read both answers in one thread and notice exactly where they disagree.

Where two strong models disagree is usually where the interesting part of the problem is. On a coding task it is often the fastest way to find the assumption one of them made silently.

It is also a cost control. Explore on an economical model, then switch up for the handful of turns that genuinely need depth:

The full catalog is on the models page, and every credit rate is in the transparency hub.

Install It in Two Minutes

1. Pick your build. Go to the download page and choose Apple silicon or Intel. If you are not sure, open the Apple menu → About This Mac and look at the Chip line — anything starting with M is Apple silicon. There are two separate disk images rather than one universal binary, and the wrong one will not launch.
2. Drag it to Applications. Open the .dmg and drag WidelAI across.
3. Launch it. The app is signed with our Apple Developer ID and notarized by Apple, so it opens normally. You should never need to right-click to bypass Gatekeeper, and if a copy of "WidelAI" asks you to, it did not come from us.
4. Open a project and send something. The picker shows what each model costs before you send.

Requires macOS 14 Sonoma or newer. No API keys to paste, and no separate desktop subscription — the plan you already have covers it.

Why It Is Not in the Mac App Store

Because it cannot be. App Store apps run in Apple's sandbox, and a sandboxed app cannot run arbitrary shell commands in your project or read files outside a narrow container. The sandbox is a good default for most apps and a hard blocker for this one.

So the app is distributed directly, as a signed and notarized disk image. That means the security work is ours to do rather than the store's, and it is worth being specific about what we do:

• Signed with an Apple Developer ID and built with the hardened runtime enabled.
• Notarized by Apple, so Gatekeeper recognises it on a Mac that has never seen it before.
• Updates are verified before they install. The app checks the code signature of a downloaded update against our Developer ID and its own bundle identifier. A build that fails is refused and your working copy is untouched.
• Nothing strips the quarantine flag. Some direct-download apps tell you to run xattr -cr to get past Gatekeeper, or do it silently during an update. That is a Gatekeeper bypass, and an app willing to do it to itself has given up the one check that would catch a tampered download.
• Updates install atomically. The new version is staged beside the old one and swapped; the previous copy is kept until the swap succeeds, so a failure mid-update leaves you with a working app.

Your Session Stays on the Mac

The Mac app stores your sign-in in the macOS Keychain, marked so it never syncs to iCloud and never migrates to another Mac. That matters more for a desktop app than it sounds: the app is unsandboxed by necessity, and an unsandboxed app's ordinary preference files are readable by anything else running as you. The Keychain is not.

The rest of our position on your content is unchanged:

• Your prompts, files, conversations, and outputs are never used to train AI models.
• Content is sent to the model provider you choose in order to generate a reply, and nothing more.
• Data is encrypted in transit with TLS/SSL and at rest using industry-standard safeguards.
• The agent reads only what it needs from the folder you opened.

Full detail is in our privacy policy.

What the Mac Is Actually Better At

• Refactors that span files. Describe the change once and review it as a set of diffs rather than pasting files into a browser one at a time.
• Reproducing the bug. The agent can run the failing test and read the real output, including the part that only happens on your machine.
• Getting into an unfamiliar repository. Point it at a project you inherited and ask what talks to what.
• The tedious pass. Renames, missing test cases, docstrings — work that is easy to specify and dull to do.
• Working offline-ish. Your history is local, so browsing past conversations does not wait on the network.

A Word on Trusting the Output

Answers are produced by AI models and can be wrong, out of date, or confidently mistaken — and on this platform they can also edit your files. Read the diffs. Keep your work in version control. Treat a green test run as evidence rather than proof. The app is intended for adults aged 18 and over, and for medical, legal, or financial questions you should check anything that matters with a professional.

What It Costs

The app is free to download. Generating responses requires an active plan, and one plan covers web, Mac, and Android:

Agent runs use more credits than chat, because each tool round trip is a request. Start on an economical model and switch up when the problem earns it. Full detail is on the pricing page.

Get It

Download WidelAI for Mac →

No account yet? Start with WidelAI on the web, then install the app and pick up where you left off. Prefer your phone? WidelAI is on Google Play too.

Related Reading

• WidelAI for Android Is Now on Google Play
• Kimi K3 vs GLM-5.3: Which Open-Weight-Era Frontier Model Should You Use?
• Claude Opus 5 vs GPT-5.6 Sol: The 2026 Frontier Coding Showdown