Winskel for agents

What Winskel is

Winskel is a multi-model coding orchestrator for macOS that coordinates coding work across Claude Code and Codex. The user gives one coding objective. Usually one strong model owns the work. When the work warrants it, Winskel plans it, runs independent parts at the same time in separate Git worktrees, and brings the work back into one review.

Winskel is an independent product. It is not made by, partnered with, or endorsed by Anthropic or OpenAI. It does not replace Claude Code or Codex; it runs the installations and accounts already on the user's Mac.

• The goal is not more agents; it is the smallest team that can finish the objective.
• There are three levels of control: Autopilot, Preferences, and Custom orchestration. Explicit developer constraints always win over automatic routing.
• Winskel prefers to keep one model on a mission unless specialization, independent review, availability, or escalation justifies a switch.
• Escalation is bounded: when a mission's checks fail, the same model gets one repair turn, then the mission moves once to a stronger model. If that also fails, the objective stops as Failed with the reason.

Who it is for

Developers on Apple Silicon Macs who already use Claude Code, and optionally Codex, and want one coding objective coordinated across them with a reviewable result. It is not for Intel Macs, Windows, or Linux, it is not a hosted cloud service, and it does not replace human code review.

Current public release

• Version: 0.7.0-alpha.1 (alpha channel). The website calls the current public stage the Beta. The downloadable build is an alpha channel build, and some screens in the app say Alpha.
• Platform: Apple Silicon, macOS 12 or later. Not supported: Intel Macs, Windows, Linux.
• Signing: The build is not signed with an Apple Developer ID and is not notarized by Apple. macOS asks the user to approve the first launch.
• Updates: No automatic updates. New builds come from winskel.com.
• Download: the Download for Mac action on the homepage. Downloading and installing needs the user's own decision.
• Installation guide: Install Winskel

Requirements

• Required: A Mac with Apple Silicon running macOS 12 or later. Intel Macs, Windows, and Linux are not supported. Details
• Required: A Google account. Winskel accounts sign in with Google. Details
• Required: Claude Code installed and signed in. Winskel plans every objective with Claude Code, so work cannot start without it. Keep Claude Code open on the Mac; Winskel detects the local app and connects automatically. Details
• Optional: Codex installed and signed in. Needed for routes that use Codex. Keep Codex open on the Mac; Winskel detects the local app and connects automatically. Details
• Required: A Git repository the user is authorized to change. A local Git repository, or one cloned from GitHub. GitHub is optional. Setting up Git for a plain folder is not in the current public download; it is in the next release. Details
• Required: An active provider plan for each agent used. An Anthropic plan for Claude Code (required), and an OpenAI plan for Codex only if Codex is used. Their subscriptions, usage limits, and charges are separate from Winskel. Details
• Required: Free disk space for Git worktrees. Each run gets its own Git worktree on the Mac. Details

With Claude Code only: yes. Winskel works with Claude Code alone and uses only Claude Code models. With Codex only: no. A Mac with Codex but no Claude Code cannot start work, because Winskel plans every objective with Claude Code.

Available capabilities

Orchestration modes

• Autopilot (Available now). Winskel makes the orchestration decisions. Nothing to configure.
• Preferences (Available now). Winskel still decides, inside boundaries the user sets: allowed and excluded models, preferred models by kind of work, and a priority.
• Custom (Available now). The user sets the implementation, review, and approval rules Winskel follows. Winskel still plans the steps and dependencies.

Capabilities

• Autopilot (Available now). The default. Give one objective and Winskel decides whether one model owns it or it becomes connected steps, what runs at the same time, and which model does each part. Limitation: Usually one strong model owns the whole objective. Winskel plans a split when the work is high risk or too large, when a step needs recovery, or when the user asks. Source
• Name a model or agent system (Available now). A model or agent system named in the objective is used for that work. Retries and escalation never replace it. If it cannot run, Winskel stops and explains instead of substituting another. Source
• Parallel work in Git worktrees (Available now). Independent parts run at the same time, each in its own Git worktree and branch. Limitation: A worktree separates Git working state. It is not an operating system sandbox. Source
• Dependencies (Available now). A step that depends on another waits until the steps it needs have finished and passed their observed checks. Source
• Needs You (Available now). Questions, permission requests, and decisions that need a person appear in one place. Limitation: Claude Code and Codex expose different question and permission capabilities. System failures show as Failed, not as Needs You. Source
• Review (Available now). One review shows changed files, the step and model that changed them, the diff, and the checks Winskel observed. Source
• Observed checks (Available now). A check counts as Passed or Failed only when Winskel has the command and its result for that run. Everything else is Not verified. Limitation: Codex runs its commands inside Codex's own sandbox, where Winskel does not run project checks, so Codex steps show as Not verified unless Codex reported a check it ran. Source
• Approve and Request changes (Available now). Approve records that the user accepts the result. Request changes sends a follow-up objective. Limitation: Approve does not merge, push, copy, or delete anything. Source
• Agent permission modes (Available now). Ask Me, Smart Auto (the default), and Autonomous decide which commands and edits an agent may run without asking. Limitation: Some actions are blocked in every mode. Agents still run under the user's macOS account. Source
• Routing reason (Available now). The model chosen for the work and a one-line reason are shown when the work starts. Source
• GitHub (Available now). Optional. Clone a repository from GitHub and open pull requests from Winskel. Source
• Preferences (Available now). Model pools, excluded models, preferred models by kind of work, and a quality, speed, cost, or balanced priority. Winskel decides inside them. Source
• Custom orchestration (Available now). The user sets the implementer, reviewer, review rules, failure handling, and approval gates. Winskel still plans the steps. Custom does not set step order, dependencies, parallelism, or a number of repair rounds. Source
• Workflow file import (Available now). Import a workflow file for Custom orchestration. Free text becomes planning guidance, never a routing rule. Source
• Routing history (Available now). A full record of which model handled which work, why, and what happened next. Source
• Apply to project and Undo apply (Next release: built and in testing, not in the current public download). Add a result to the project folder from Winskel, and take it back out. Source
• Git setup for a plain folder (Next release: built and in testing, not in the current public download). Start a project from a folder that is not yet a Git repository. Source
• Free plan limits and Pro checkout (Available now). Free limits are enforced, and Pro checkout starts in the Winskel app at Settings → Billing. Source
• Routing that learns from outcomes (Not available). Routing follows a fixed, versioned policy. It does not learn from the user's code, history, reviews, or ratings today. Source
• Winskel Cloud (Not available). Not part of the current product. Source
• Remote phone control (Not available). Not part of the current product. Source

How the workflow operates

1. The user adds and trusts a Git repository, or clones one from GitHub.
2. The user writes one coding objective.
3. Winskel decides whether one model owns the work or the objective becomes connected missions. Usually one strong model owns it.
4. Winskel gives each mission to a model the user's Claude Code or Codex can run. A model or agent system the user named is used instead.
5. Independent missions run at the same time, each in its own Git worktree and branch. A mission that depends on another waits for it.
6. Questions, permission requests, and decisions that need a person appear in Needs You.
7. Review shows the changed files, the mission and model behind them, diffs, and the checks Winskel observed.
8. The user approves or requests changes, then brings the result into the project with Git or a pull request.

Supported models

• Claude Code: Fable 5.1, Fable 5, Opus 5, Opus 5.5, Sonnet 5, Sonnet 5.5
• Codex: GPT-6 Astra, GPT-5.6 Sol, GPT-5.6 Terra, GPT-6.1 Sol, GPT-6 Sol
• Default models Winskel routes work to, in order: Sonnet 5.5, Opus 5.5, GPT-5.6 Sol. Winskel tries these in order and uses the first one the user's Mac can run, unless the work calls for something else or the user names a model.
• A model is used only when Winskel supports it, the installed Claude Code or Codex reports it can run it, and the user's provider account allows it. A Mac can show fewer models than this list.
• Not presented as workers: Cursor, Gemini, Grok, Copilot, OpenCode.

Supported models, with model ids

Pricing and limits

• Winskel Free: $0 a month (USD). 5 active projects, 5 active routing models, 150 routes a month. No payment. Downloading Winskel and signing in uses Free.
• Winskel Pro: $19 a month (USD). No plan limits. Upgrade inside the Winskel app at Settings → Billing. The website does not sell Pro directly.
• Enforcement: Free limits are enforced. Winskel warns at 120 and 140 routes in a month; route 150 succeeds, and the next route waits until the monthly allowance resets or the user upgrades. Running work continues. Pro checkout starts in the Winskel app at Settings → Billing.
• Route: A route is one executable mission Winskel sends to a coding model. Planning and Winskel's own retries on the same model do not use additional routes; a rerun the user starts is new work.
• Active routing models: Winskel detects and shows every supported model. On Free, up to 5 are active in your routing pool at one time.
• Winskel's price does not include Claude Code or Codex. Anthropic and OpenAI subscriptions, usage limits, and charges are separate and still apply.

Pricing

Security and data handling

• Claude Code and Codex run on the user's Mac, under the user's macOS account and provider accounts.
• Leaves the Mac: Claude Code and Codex send prompts, repository content, and other task information to Anthropic and OpenAI under the user's own agreements with them.
• Leaves the Mac: Winskel's hosted service stores the records it needs to coordinate and explain work.
• Leaves the Mac: Google sign in shares the user's name, email address, and profile picture with Winskel.
• Stored by Winskel: Email, project names and paths, objective and chat text, plans and answers, notes, question and permission text, run status, bounded run events (agent messages, command lines, and short tool output excerpts, which can include code an agent read), summaries, changed file paths and line counts, branch names and commits, attachment names and paths, verification evidence, routing decisions, and provider usage readings.
• Not stored: Repository files, patches, and attachment files are read from the Mac and are not stored by Winskel; only the short excerpts in run records can contain code. Winskel does not collect or store Claude Code or Codex credentials.
• Local execution does not mean nothing leaves the Mac. Records and short excerpts can contain source code, paths, command output, or secrets.
• Website: The public website uses no advertising trackers or third-party cookies. It uses Vercel Web Analytics to count anonymous page views in aggregate, and Vercel processes normal request information to deliver the site.
• Winskel starts work only in repositories the user explicitly adds and trusts.
• Agent permission modes (Ask Me, Smart Auto, Autonomous) control which commands and edits an agent may run without asking. Some actions, such as sudo, git reset --hard, force push, and paths outside the trusted worktree, are blocked in every mode. These are permission rules Winskel applies to agent requests, not operating system enforcement.
• Agents run under the user's macOS account. Git worktrees separate working state but are not an operating system sandbox.
• Approve a first launch only for a build downloaded from winskel.com.

Local execution · Privacy · Security

Known limitations

• The build is not signed with an Apple Developer ID and is not notarized by Apple. macOS asks the user to approve the first launch.
• Apple Silicon Macs with macOS 12 or later only. No Intel Macs, Windows, or Linux.
• No automatic updates. New builds come from winskel.com.
• Some screens in the app say Alpha. They belong to the same public build.
• Claude Code is required. Codex is optional.
• Free limits are enforced. Winskel warns at 120 and 140 routes in a month; route 150 succeeds, and the next route waits until the monthly allowance resets or the user upgrades. Running work continues. Pro checkout starts in the Winskel app at Settings → Billing.
• Routing does not learn from the user's feedback or history yet.
• A numbered list in an objective runs as one step unless the user asks for separate or parallel steps.
• Codex steps show as Not verified unless Codex reported a check it ran.
• Revert, which adds an inverse commit for one objective's result when Winskel can prove it is safe, is in testing.
• Provider usage readings update only when that agent runs and can be stale.
• Winskel Cloud and remote phone control are not available.
• Provider limits and model availability still apply.
• Git worktrees are not a filesystem sandbox.

Known limitations

How to read the evidence on winskel.com

• Current capabilities: What the current public download does today. This reference marks each one as available.
• Next release: Built and in testing in the release candidate, not in the public download yet. Screenshots of these features are captures of the release candidate.
• Workflow illustrations: Diagrams labeled as illustrations, built from Winskel's example project. They explain the workflow; they are not recorded runs and contain no measured times, token counts, costs, or percentages.
• Winskel Lab method and design: The Winskel Lab page (https://www.winskel.com/lab) explains how Winskel intends to compare models and workflows. Its workflow explorer, quality and cost diagram, and experiment matrix are conceptual and labeled as such: they contain no data points, scores, costs, or percentages, and each matrix cell is a proposed comparison.
• Measured comparisons: None are published yet. Winskel does not currently claim a measured speedup, a token saving, or a quality improvement.
• Research questions: Open questions on the Winskel Lab page, such as when a second model reviewing the work catches problems that a second pass by the same model misses. They are not findings.
• Long-term vision: Winskel intends to study research and data analysis workflows. No product for those fields exists or is available, and no launch date is set. These are directions, not products.

Availability is stated in text. A screenshot alone does not show whether a feature is in the public download.

Winskel Lab research

Winskel Lab is Winskel's research program for comparing AI models alone and in workflows on specific tasks, starting with coding.

• Purpose: Compare individual models and model workflows on specific tasks, to find when a combination improves the result, when it reaches comparable quality at lower cost, and when one model is enough.
• Current focus: Coding. Future directions: Research workflows and Data analysis workflows. Directions to study later, not products. No launch date is set.
• Method: Whole configurations are compared (model version, role, tools, context, settings, retries, verification and workflow) under equal conditions, against a well configured single model, with independent evaluation and confirmation on separate tasks. Failures stay in the denominator, and unknown values are reported as unknown, never as zero.
• Coverage: Winskel aims to evaluate models from multiple providers. Planned coverage is not evaluated coverage. Evaluated: None published. Winskel does not claim a count of evaluated models or providers.
• Findings: None published yet.
• Product boundary: Lab research is separate from the product. A model evaluated in the Lab is not thereby supported by the product, and Lab results do not change the product automatically: routing follows a fixed, versioned policy, and learning from routing outcomes is not available.

Winskel Lab

A few things to know

Terms

Objective

The outcome the user asks for in one message, and the record of what happened to it.

Mission

One unit of work within an objective, given to one agent and one model. Most objectives are one mission. The app also calls a mission a step.

Agent

A coding agent tool Winskel runs on the Mac: Claude Code or Codex.

Model

The language model an agent runs for a mission, such as Sonnet 5.5 or GPT-5.6 Sol.

Provider or harness

The agent system and the company behind it: Claude Code from Anthropic, or Codex from OpenAI. Winskel uses the user's own installation and account.

Working copy

The separate Git worktree each run gets, on its mission's own branch, so concurrent work never shares one directory.

Route

A route is one executable mission Winskel sends to a coding model. It is also the unit the Free plan counts.

Review

Where the user inspects the result: changed files, the step and model behind them, diffs, observed checks, and the decision to Approve or Request changes.

Apply

Adding a result to the project folder from Winskel (Apply to project) and taking it back out (Undo apply). Next release: built and in testing, not in the current public download. In the current public build the result stays on its own Git branch, and the user brings it into the project with Git or a pull request.

Verification

The evidence that a check ran: Passed or Failed only when Winskel has the command and its result. Not verified means Winskel did not receive sufficient evidence that a check ran. An objective as a whole is never labeled verified.

Needs You

The list of questions, permission requests, and decisions that need the user. A system failure is shown as Failed, not as Needs You.

Important differences

Common questions

Setup problems

• “Winskel needs Claude Code on your Mac. Install it, then run claude auth login in a terminal. Codex is optional.” Fix: Install Claude Code, then run claude auth login in a terminal.
• “Codex is available, but Winskel plans every objective with Claude Code, so work cannot start until Claude Code is signed in.” Fix: Run claude auth login in a terminal.
• “Not signed in. Run codex login in a terminal on this machine.” Fix: Only needed to use Codex models. Run codex login in a terminal.
• “Winskel cannot start work because the runner on this Mac is not running. Start it again, or quit Winskel and open it again.” Fix: Quit Winskel and open it again. If it persists, open Diagnostics in the app.
• “macOS cannot verify that Winskel is free of malware (first launch).” Fix: Only for a build downloaded from winskel.com: System Settings, Privacy & Security, Open Anyway.

Where to find detailed instructions

• Install Winskel (Markdown): Download Winskel for Mac, approve its first launch, and sign in with Google.
• Connect Claude Code (Markdown): Claude Code is required. Winskel uses the Claude Code installation and account already on your Mac.
• Connect Codex (Markdown): Codex is optional. Winskel runs Codex under your existing account when it is installed and signed in.
• Your first objective (Markdown): Start with an outcome that can be checked, not a list of agents to operate.
• Choosing models (Markdown): Winskel decides by default. Preferences set the boundaries, and Custom sets model roles, review rules, and approval requirements.
• Model routing (Markdown): One routing and policy system decides which model does the work: your constraints first, then what your accounts can run, then the work itself.
• How Winskel breaks down work (Markdown): Most objectives stay with one model. Larger or riskier ones become a small set of connected steps that Winskel runs and shows you.
• Projects and worktrees (Markdown): Projects point to trusted Git repositories. Each run receives its own Git worktree.
• Needs You (Markdown): Needs You groups real decisions and permissions that require a person.
• Review and verification (Markdown): Review separates what changed from what was actually checked, and records your decision.
• Permission modes (Markdown): Choose how often agents ask before commands and edits. This is separate from how Winskel routes work.
• Local execution (Markdown): Coding agents run on your Mac in repositories you choose. Your providers still receive what their tools send.
• Supported models (Markdown): A model is eligible only when Winskel, the installed provider tool, and your account all agree.
• Known limitations (Markdown): The current Beta is local, macOS only, Apple Silicon only, and intentionally narrow.

• Winskel: What Winskel is, the demo, and the download.
• Product: How Winskel coordinates Claude Code and Codex, in one page.
• Orchestration: Autopilot, Preferences, and Custom, and how a named model wins.
• Parallel work: Parallel missions, dependencies, and Git worktrees.
• Review and checks: What changed, which checks were observed, and the decision.
• Supported models: The Claude Code and Codex models Winskel can route to.
• Pricing: Free and Pro, what a route is, and what is enforced today.
• Documentation: Setup, concepts, security, and troubleshooting.
• Changelog: Dated public product updates.
• Security: Trust boundaries and vulnerability reporting.
• Privacy: What Winskel collects and why.
• Winskel Lab: Research on comparing AI models alone and in workflows on specific tasks, starting with coding.
• Manifesto: Why Winskel exists.
• Contact: How to reach Winskel.

Not confirmed

• New account sign ups: Sign in uses Google. Whether anyone can create an account today, or only invited users, was not confirmed in this review.

Support: ahmed@winskel.com · Contact · Security