How Winskel works

Four separate parts

Winskel coordinates coding agents. It does not replace Claude Code or Codex, and it does not run a model of its own. Four parts take part in every piece of work, and they stay separate:

Part	What it is	What it does
The Winskel Mac app	The app you install, with a local runtime it starts and stops	Shows the conversation, lets you choose and trust project folders, and runs the work on this Mac
Winskel's service	Your Winskel account, hosted by Winskel	Keeps your objectives, plans, project memory and the records of what ran, and decides what runs next
Claude Code and Codex	The command line tools installed on your Mac	Do the coding under your own sign-in, in a working copy Winskel prepares
Anthropic and OpenAI	The providers behind those tools	Receive what Claude Code and Codex send them, under your accounts and their terms

Winskel's service does not run the coding agents. Claude Code and Codex run on your Mac, and they are the part that talks to Anthropic and OpenAI.

What runs where

• On your Mac: Claude Code and Codex, the Git working copies they write in, the checks Winskel runs, Preview and Terminal. File contents and patches are read on your Mac.
• In Winskel's service: your account, projects, conversations, plans, project memory, and the records that explain each result, such as which model ran, the commands Winskel observed, and the checks and their results. Local execution has the full list.
• Only the Mac app can trust a project folder and start agents in it. A web page cannot do either on its own.
• Your Claude Code and Codex sign-ins stay with those tools. Winskel checks that they are signed in, and does not collect or store their credentials.

From Send to review

What happens to one message in Build:

1. You write the outcome you want. The Interaction menu says what the message is for (Auto, Chat, Plan or Build), and the Model menu says who may do it (Auto, Claude Code, Codex, or one exact model).
2. Winskel decides which model owns the work, inside your choice and your Preferences, and shows the model and a short reason. See how Winskel picks a model.
3. Your Mac takes the work, creates a separate Git working copy and branch for it, and starts Claude Code or Codex there with your request and a short, labeled brief from the project's memory. See shared memory.
4. The agent works on its own inside Winskel's permission policy. Anything outside it, such as deleting files, pushing or reading secrets, waits for you in Needs You. See agent permissions.
5. Winskel runs the project's checks in its check sandbox and records only what it saw run: the command, its exit code and the test counts. An agent writing that tests pass is not counted.
6. Review shows the changes, the checks and a receipt of what ran. Your project folder changes only when you choose Apply, and Git stays yours.

One model, or several when you ask

Send gives the whole objective to one model, including its repairs and any plan-first or review pass. Orchestrate, only when you press it and confirm, plans the objective, runs independent missions at the same time in separate working copies, and brings them back to one review. See Send and Orchestrate.

Controls, not guarantees

• A project folder must be trusted on your Mac before agents start there.
• Each run has its own Git working copy and branch, so runs do not overwrite each other. A working copy separates Git state; it is not a security sandbox.
• Project checks, and scripts an agent wrote, run inside a sandbox with the network off and no access to your other files.
• One permission policy decides which actions run without asking and which wait for you.
• Winskel refuses a run that would bill your provider per use, unless you allow it on that Mac. See tokens and plan usage.
• Claude Code and Codex still run under your macOS account and send prompts and code to Anthropic and OpenAI under your accounts. Local execution does not mean nothing leaves your Mac.