A frontier grade harness for open models.

Cowork is the agent for documents and folders. Give it a goal in the composer and it plans the work and does it inside the folders you grant. It asks one question at most, and only when it must, and it checks that the file it says it wrote is really on disk before it says done.

Ask it to write something and the full text comes back in the chat, and is also saved as a Word (.docx) document. It makes decks, PDFs, spreadsheets, documents and working HTML. Every output keeps every version, and every run can be undone from the conversation. The walkthrough above is a scripted demo: Cowork builds a Flappy Bird game, Code fixes it, and Cowork writes an essay as a PDF and makes a deck.

Code builds and runs projects, and each project lives in its own project folder. It tests its work, and ▶ Present opens what it built so you can use it, for example by playing the game it just made.

It has a built-in terminal, and background tasks for things like dev servers, which you can stop. Helper agents take parallel edits and are shown live in the conversation, and a live working bar shows what it is doing, how long it has been at it and the tokens it is using, as it streams.

Chat is the conversation. The built-in model, Aletha Core, runs on your Mac and works on first launch, with no account, or pick any model from Ollama Cloud or your local Ollama; the model it is talking to is named above the composer. It uses the same memory and the same permissions as Cowork.

Voice runs on your Mac. The floating mic is a small round button that stays over whatever you are in; tap it and it listens, thinks and answers out loud in the Kokoro voice. It listens only when tapped. It can be dragged anywhere on the screen, collapsed back to the button, or hidden, and its Chat toggle opens a typed version of the same thing.

Aletha keeps a memory layer of its own, separate from whatever model is answering. When something useful comes up, such as a name, a preference, a project or a decision, it saves it with where it came from and why it kept it. The Memory page lists every entry: search them, pin the ones that matter, delete any of them, or open the conversation an entry came from.

Because it lives outside the model, it carries across conversations and across model swaps. Switch from the built-in model to a 30B or a model through your Ollama account and you do not repeat yourself. Near-duplicates are merged, a contradiction keeps the newer fact and archives the older one, and memories you never use fade rather than being deleted.

TVs and speakers, by voice or by text, over your own network. Say “play something in the living room”, “volume down” or “quiet everything”, from Main, from Cowork, or from the floating mic. Roku TVs today. The Devices page is Device, Connected Devices and Discover Devices: Scan for TVs, Scan for speakers, Add TV / media, and Make this the main TV.

Title lookups (“put Rick and Morty on”) go through a free Watchmode key you paste into Settings › Devices; it is kept in your Mac's keychain, and nothing else about your devices leaves the Mac. Without the key it still opens apps and controls the TV. One step is on the TV itself: Roku's Network access set to Permissive (some call it Enabled). The app cannot change that from the Mac.

A calendar of your own, with Month, Week and Day views and a Calendar Assistant beside them that you type or speak to: “Add meeting July 5th at 3 PM”, what's on today, move something. Events can also be added, edited and removed by hand. It is stored locally on this device; the page says so at the top.

Scheduled tasks can read it: a Monday brief of the week's meetings, written into a note, is one of the presets.

Goals that run on a cadence while you are away: Every day, Weekdays, Every Monday, First of the month. Each has a name and its own instructions, and a Run now for trying one immediately. Results land in the same Outputs as any other run, on the same log. If the agent needs a decision from you while you are away, the task waits, and the Scheduled page shows it under “While you were away” when you return.

Two limits the app states plainly. Scheduled tasks use the folders you have already given Aletha access to; they cannot be given a folder of their own, and they cannot approve anything while you are away. And if your Mac is asleep or Aletha is closed when a task is due, it will not run at that moment.

A private workspace behind its own passphrase, separate from the rest of the app. The first time you open it you choose and confirm a passphrase and create the vault; after that it asks for the passphrase to unlock. What you keep inside is AES-256 encrypted on disk, as the vault's own top bar says, and the model does not read it: nothing in the vault is sent to a model, local or cloud.

It is a place for the things you would not paste into a chat.

Every run works under one of three approval modes, chosen from the pill above the composer, and the pill always shows what is being enforced right now. Ask every time: nothing runs until you approve it. Ask before changes: reads in folders you've allowed run on their own; anything that changes, runs, installs, or leaves your Mac asks. Skip permissions: nothing asks, including deletes, shell and network. In Skip permissions, test runs stay in a sandbox with no network access.

Three things still stop in every mode: a capability you turned off, a target outside your folders, and a missing credential. Deletes stay recoverable from their receipt. Every action is written to the log with who allowed it, and a step the mode let through is marked “ran without asking” in the transcript.

Tested on kimi-k2.7-code and gpt-oss:120b through Ollama Cloud, and works with any model you run in Ollama. Local models keep everything on your Mac; cloud models send prompts to the provider you chose, and the app says so under the composer.

The built-in model, Aletha Core, runs locally and works on first launch with no account. Multi-step work like a game or a deck wants a 30B-class model or a cloud one, through Ollama Cloud or a local Ollama.

What leaves the Mac: your prompts, to a cloud model, while one is selected; title lookups to Watchmode, if you set a key. Nothing goes to Aletha: no account, no telemetry.

Aletha Codex 2.1.1 needs a Mac with Apple Silicon (Intel Macs are not supported), macOS 12 Monterey or later, 8 GB of RAM and 11 GB free; the download is 3.6 GB. For any model beyond the built-in one you need Ollama Cloud or a local Ollama. The built-in model, Aletha Core, runs locally with nothing else on the machine. No account and no sign-up: you pay once, the license key arrives on the next page, and the app activates with it. If it isn't for you, email within 30 days for a full refund, no questions asked.

Mac only right now. Windows or Linux: join the waitlist. How many people ask decides the order.