Wayland AI 工作流

Cowork: plan, build, and ship anything from one place

AI finally gets useful when it stops answering one question at a time and starts doing the whole job. That is Cowork: you and the system work the problem together, from the first rough idea to the shipped result. Write the book. Run the multi-week project. Plan, build, and ship the thing you actually came to do.

It is collaborative and autonomous at the same time. You set the goal and stay in control. Wayland keeps state across the whole arc of the work: your files, the running plan, and a memory that does not reset between sessions. When a job is bigger than one chat, it scales up on its own:

• One assistant, real file access. Cowork operates directly on your files with the built-in tools (Read, Write, Edit, Bash, Grep, Glob), processes real documents, and plans multi-step work. Start in a read-only Plan mode to see the approach before anything is touched.
• A self-assembling team when the job is big. Hand the goal to a team. A manager you talk to in plain language assembles the right specialists and they work the task together against one shared blackboard until it is done. Keep a Standing Company on call, or spin up an ad-hoc team for a single push.
• Pre-trained specialists on tap. Pull in a domain expert for the part that needs one, then drop back to driving the whole thing yourself.
• It keeps going on a schedule. Put the recurring parts on cron and they run while you are away.

You are not prompting a chatbot. You are coworking with a system that remembers the project, owns the grunt work, and checks in when it matters.

Build from source

Requirements: Bun 1.3.x, Node 22 to 24, and your platform toolchain for native modules.

```bash git clone https://github.com/ferroxlabs/wayland.git cd wayland bun install

Or run it with no install

npx @ferroxlabs/wayland-core "Read Cargo.toml and explain the dependencies" ```

The CLI self-updates with npm update -g @ferroxlabs/wayland-core, independent of the desktop app.

Configuration and keys

Wayland runs on your provider credentials. There is no required Wayland-hosted backend to chat.

Engine key resolution order: --api-key, then config, then API_KEY env, then provider-specific env.

Workflows from idea to outcome

178 ready-to-run workflows that take you from a blank page to a finished deliverable: a launch plan, a competitor teardown, a month of content, a release write-up. Run one as-is, build your own, or put it on a schedule. Each one walks the steps so you get the outcome, not just a starting point.

<p><img src=".github/assets/workflows.jpg" alt="Wayland workflow running step by step toward a finished outcome" width="100%"/></p>

Approval modes for scheduled tasks and workflows

New scheduled tasks and workflow runs use guarded approval defaults: Auto-Accept Edits for Wayland Core and Gemini, Accept Edits for Claude, and Auto Edit for Codex. Other backends use their normal mode without Desktop enabling blanket auto-approval. Choose the approval mode in the scheduled-task editor or the workflow launch dialog; a global preference from another conversation does not authorize a new workflow to run in Autopilot.

For Core, Auto-Accept Edits allows the engine’s built-in Write/Edit behavior and its configured auto-approved tools. Other operations can require a confirmation. Approval mode does not widen workspace filesystem boundaries or disable Core’s allow-listed network tools. Select Autopilot or Full Auto explicitly for a run only when that is the intended policy. Existing tasks retain supported stored modes when edited; review that field when updating an older task. Unrecognized Core mode values, including the old generated bypassPermissions token, use ask-first rather than enabling full-auto. Newly seeded Core routines use Auto-Accept Edits.

A scheduled turn is not dispatched when its requested settings still fail after a fresh-session retry. Starting a fresh workflow uses the displayed approval choice; resuming an existing workflow returns to that conversation’s current policy.

FAQ

Are my keys and data private? Yes. Keys are stored in the OS keychain and data lives in SQLite on your disk. There is no Wayland account and no Wayland server: when you send a prompt it goes to the provider you configured, on your key, and nowhere else. The engine's outbound traffic passes an egress gate that is enabled by default and classifies every request before it leaves. "Air-gapped" would be the wrong word for a tool you point at a cloud model, and it is the provider you chose that sees your prompt.

Can I run fully offline? Yes. Point the engine at a local Ollama model and Wayland runs with no network at all. Voice dictation runs offline with a bundled Whisper model, and Wayland can read replies back to you with voice output, so you can work hands-free.

Do I need a Wayland account or subscription? No. Bring your own provider keys, or use the Claude Code backend with your Claude subscription (sign in with the claude CLI). There is no paywall to do work.

My local models are not showing up. Make sure Ollama is running and reachable on its default port, then refresh the model list in Settings.

Is it really open source? Yes, both parts. The desktop app is GNU AGPL-3.0 and the Wayland-Core engine is Apache-2.0, so you can embed the engine anywhere.