Workers for the folders you work in.
Download for Mac · Website · Contribute · Docs
work-fold is an open-source Mac app that brings your files, AI conversations, and project tools together. Open a folder, tell its Worker what you need, and work on the actual files. Research, writing, planning, code—whatever belongs in that folder.
For example: open a folder of vendor quotes, ask for a comparison, and have the Worker write a decision brief beside the originals. Come back tomorrow and pick up the same conversation.
Each work-folder has its own Worker and conversations. Your files stay ordinary files, accessible in Finder and your existing tools. A work-folder is an ordinary folder you choose to work in; it is never converted into a proprietary container.
The work-fold agent is the agent above your work-folders. Ask it what needs attention or have it coordinate work between projects, from the app or Mac menu bar. Hand a piece of work to one work-folder's Worker and it reports back here; if it needs something first it asks, and your answer picks the work up where it stopped.
You can register work-folders inside another work-folder. The switcher shows
their hierarchy, and Files opens each nested work-folder in its own context.
Type @ in a Chat or the work-fold agent's composer to address a Worker.
The local work-fold work-folders list --json projection includes parentWorkFolderId
for a nested work-folder; see the work-fold agent and CLI guide.
As a project grows, ask for a custom app in its sidebar, add Checks to review chosen files, or set up an Automation to run fixed steps on a schedule or after selected files change. Automations are deterministic app-run steps, not a second kind of Worker or an open-ended agent scheduler. These are optional; start with a work-folder and a conversation.
The built-in browser shares its tabs with the work-fold agent and every Worker. They can read and use pages, inspect screenshots, hover and drag, work inside scrolling panels, attach local files to file inputs, and find completed downloads. Its sign-ins belong to work-fold's own profile, separate from Chrome.
What a Worker or the work-fold agent does happens right away and leaves a record you can read in the app. Nothing it deletes is gone for good: History keeps versions of your files, and anything History cannot keep waits in Recently deleted for 30 days. An app's access and anything running on a schedule can both be turned off afterwards.
Use Compare in Version History or a selected work-folder restore point to read saved text beside the current file before restoring. Binary, oversized, uncaptured, and incomplete comparisons explain their limits. The authenticated CLI offers paged history list/versions, verified history read --offset-bytes 0 ranges for large saved text, and streaming search with --path narrowing and --cursor continuation; see work-fold help history and work-fold help search.
- Download work-fold for an Apple silicon Mac.
- Open it. First run connects your AI (sign in with ChatGPT, Claude, or OpenRouter, or paste an API key; provider usage may cost money), optionally sets up Chrome and Computer Control, then asks for your first folder. You can change the model anytime in Settings → AI Models.
- Choose Set up for me to have the Worker organize the folder and give it a look, or just start a Chat.
Microsoft Foundry chat deployments connect in Settings → AI Models using separate Responses, Chat Completions, and Anthropic connections. Enter the endpoint, API key, and deployment names for each format you use; all three can stay connected together. Existing Azure connections continue to work. See model connections for setup and scope.
Files live on your computer. Content a Worker uses goes to your chosen model provider. See Privacy for details.
Handed-off work is tracked as a durable request: a question waits for its answer and continues once, and each Worker reports one result to whoever asked. The request graph and original assignments stay machine-local; only results a Worker deliberately releases reach another work-folder's Chat, and an app's task reads stay pinned to its own installation. See the collaboration contract for stop, expiry, and restart behavior, and the collaboration experience for how progress, questions, and result files appear in Chats, the work-fold agent, and apps.
Claude Code, Codex, and other AI agents can use work-fold too: Settings →
Command Line adds the work-fold command to Terminal and adds a short skill
that teaches an agent to hand work to your Workers, follow it, and run
Automations. See the CLI guide.
Long work runs in the background: a Worker or the work-fold agent starts an export, build, install, or server as a background job, keeps working or ends its turn, and is brought back to the same Chat once with the result when the command finishes. History captures what the job changed, its output stays on this computer, and Stop or quitting ends it. See Background jobs.
Workers also have structured collaboration tools, asynchronous questions,
interactive terminal input, managed Git working copies, and focused read-only
helpers. Workers and the work-fold agent can recover completed tool evidence
and saved findings after compaction with progress, including file freshness
checks. See Worker tools for their lifecycle and limits.
The work-fold agent's menu-bar popover has a Chats view that reopens saved conversations and keeps unsent drafts while you switch.
Skills & Extensions manages native Pi resources, package updates and removal, and setup for the included Computer Control, Chrome, Web, Documents, and Service Connections tools, which use the same Pi formats and tool loop as Extensions you add. Computer permissions, the Chrome companion (the work-fold extension in the Chrome Web Store), and service credentials are explicit setup; web search and the document runtime work without them. An ordinary Extension's questions appear in its owning Chat and end with Stop or the session; a Worker's durable questions keep their own continuation contract. See Extensions and computer work for supported jobs, dependencies, reviewed patches, and release acceptance.
Developers can inspect upcoming model calls through the local ?dev-context
diagnostic route. Recording starts off, has no Chat or Settings entry, and
keeps assembled context, provider payloads, and source provenance in bounded
local memory. The feedback contract keeps Pi's tool
loop for text, data, and selected images without continuous screen recording.
This is an actively developed independent project, and I'd love people to build and maintain it with me. UI polish, reliable tests, clearer onboarding, and real-world bug reports would all help.
Open an issue with something you'd like to improve, or start with the contributing guide. Small, focused contributions are welcome.
Built with Electron, React, TypeScript, and the Pi agent runtime. To run a clone locally, use Node 24 (.nvmrc):
npm ci
npm run local:devOpen http://localhost:5173 for the browser development UI. See Contributing for native Electron and agent setup.
Check changes with npm run check and npm test; run npm run desktop:prepare for desktop integration changes. The docs map covers architecture, product decisions, and release procedures. AGENTS.md is the shared contributor contract.