AutoOffice 1.0.1

AutoOffice 1.0.1 — macOS and Linux, and the AutoOffice name.

A fully staffed office of AI agents, running on your own machine. This release adds macOS and Linux builds and renames the shipped application from AgentOffice to AutoOffice.

Downloads

Platform	Archive	Size	Run
Windows x64	AutoOffice-1.0.1-win32-x64-part1-app.zip and part2-training.zip — both required	773 MB + 2.0 GB	Extract both into the same folder, then run AutoOffice.exe
macOS, Apple silicon	AutoOffice-1.0.1-darwin-arm64.tar.gz	214 MB	Extract, then open AutoOffice.app — see the macOS note
macOS, Intel	AutoOffice-1.0.1-darwin-x64.tar.gz	227 MB	Extract, then open AutoOffice.app — see the macOS note
Linux x64	AutoOffice-1.0.1-linux-x64.tar.gz	235 MB	Extract, then run ./AutoOffice — see the Linux note

Windows ships as two archives only because its package (3.3 GB, most of it the offline training runtime) is above GitHub's 2 GiB per-file limit. Choose Open workspace on first launch and pick a folder; the office is created inside it.

What each platform ships

	Windows x64	Linux x64	macOS (both)
The application, Codex, local inference	yes	yes (CPU)	yes (Metal on Apple silicon, CPU on Intel)
CUDA local inference	yes	no	no
QLoRA model training	yes	no — shown as unavailable in Administration	no — shown as unavailable
In-app update staging and activation	yes	no — reported as Windows-only	no — reported as Windows-only

macOS note

The bundles are ad-hoc signed, not notarized (there is no Apple Developer ID on this build), so Gatekeeper refuses the first launch. Either right-click AutoOffice.app → Open → Open, or clear the quarantine attribute: xattr -dr com.apple.quarantine AutoOffice.app.

Linux note

Electron's sandbox needs unprivileged user namespaces. Ubuntu 23.10 and later restrict them for applications installed outside the package manager, which shows as The SUID sandbox helper binary was found, but is not configured correctly. Either allow the helper once — sudo chown root:root chrome-sandbox && sudo chmod 4755 chrome-sandbox inside the extracted folder — or start with ./AutoOffice --no-sandbox (weaker renderer isolation). Distributions that allow user namespaces need nothing.

Verification status — read this before relying on a platform

• Windows: full unit suite (1,186 tests), packaged application booted and exercised, existing user data carried across the rename.
• Linux: the full unit suite — including the file-confinement conformance tests — passes natively on Linux against the new helper (1,173 tests); the packaged application was launched under WSL2 / Ubuntu 24.04 and seeded an office. The confinement helper also passed an independent adversarial review (symlink escapes at every path segment and operation, a 7,000-iteration live parent-swap race, ASan/UBSan) with no escape found.
• macOS: cross-built and cross-signed from Linux. Fuses verified and signatures verified, but not executed on macOS hardware. The confinement helper runs its own adversarial self-test every time the application starts and refuses to enable project file tools if that fails, so a platform-specific defect fails closed rather than weakening confinement. Treat macOS as a candidate build until someone runs it.

Upgrading from 1.0.0

The application name changed, so its per-user data folder moved. On first start, 1.0.1 renames the existing AgentOffice data folder to AutoOffice — atomically, before reading it — keeping your Codex sign-in, credentials, and local runtimes. Office workspace folders are unaffected.

Known limitations

• A hard link created inside a project root that points at a file outside it can be read through the file tools. This is identical on the Windows helper (unchanged since 1.0.0), requires write access inside the root already, and Linux's default fs.protected_hardlinks=1 prevents linking files you cannot already read.
• File names containing a backslash are legal on macOS and Linux but are rejected by the project file tools, which enforce Windows-portable names in every project.

Checksums (SHA-256)

a9ed7b05e6f676195cd31f1e94f739e013094f330589ab81d71c0a3980241f21  AutoOffice-1.0.1-win32-x64-part1-app.zip
7ce1db778cd28d3261e5d6c5710ffe8ab5cda737284cc6d2b9448b2518e35c47  AutoOffice-1.0.1-win32-x64-part2-training.zip
2e206574a952bb31a552140644e28e3e4f5cbb35034f855dde51d023df1ebab3  AutoOffice-1.0.1-darwin-arm64.tar.gz
387328bab6f15747884994209f554434fa2a58b7934d07d8d5744e90a2120bf7  AutoOffice-1.0.1-darwin-x64.tar.gz
823b9465996e5ff3b63fcdd292a72f2a1f5d8f8470c2f84cc871261aed480ef5  AutoOffice-1.0.1-linux-x64.tar.gz

License

Apache License 2.0. Third-party components carry their own licenses; the Windows CUDA path includes the NVIDIA CUDA runtime under the NVIDIA CUDA Toolkit EULA. Models, adapters, and training sources you import remain under their publishers' licenses.

AutoOffice 1.0.0

AutoOffice 1.0.0 — a fully staffed office of AI agents, running on your own machine.

Each agent has a role, a private memory, and a place in the chain of approval. Everything the office decides and produces is recorded in one SQLite database inside a workspace folder you own. A new office opens with 17 divisions, 54 teams, and 179 identities.

Download

This release is Windows x64 only. The application's path-confinement, shell, and update helpers are native Windows components, and the bundled Electron, Node.js, llama.cpp, and Codex runtimes are Windows builds. There is no macOS or Linux build.

The package is 3.3 GB, above GitHub's 2 GiB per-file limit, so it ships as two archives. Both are required.

Archive	Contents	Size
AutoOffice-1.0.0-win32-x64-part1-app.zip	The application, local inference (llama.cpp CPU and CUDA), Codex	773 MB
AutoOffice-1.0.0-win32-x64-part2-training.zip	The offline Python/CUDA training runtime	2.0 GB

1. Download both archives.
2. Extract both into the same folder. They merge into one AutoOffice-1.0.0-win32-x64 directory.
3. Run AgentOffice.exe inside it. (The executable carries the application's internal name.)
4. Choose Open workspace and pick a folder; the office is created inside it.

Requirements

• Windows 10 or later, x64
• An NVIDIA GPU is optional. CUDA-accelerated local inference and QLoRA training use it; a CPU build of llama.cpp is included and works without one.

Checksums (SHA-256)

aa779f21a5f907a495f1cefe20029003248e55b77c9a5ae28737ccaa13ce452c  AutoOffice-1.0.0-win32-x64-part1-app.zip
ee978bdc5c1c30298c94f6a55bf637de8f3cc38a44ff0e8c855b73a0449b099f  AutoOffice-1.0.0-win32-x64-part2-training.zip

License

AutoOffice is licensed under the Apache License 2.0. It bundles third-party components under their own licenses. The CUDA-accelerated inference path includes the NVIDIA CUDA runtime, which is governed by the NVIDIA CUDA Toolkit EULA rather than an open source license; the CPU-only path carries no such restriction. Models, LoRA adapters, and training sources you import at runtime are not distributed with this software and remain subject to their own publishers' licenses.