OpenZero 7.3 — independent, self-hosted AI nodes

Static project home · Download 7.3 · Preserved ZMath app

The project home is free static hosting. It is not an inference server or a required node; your runtime operates independently after setup.

OpenZero combines local model chat, a browser workspace, an OpenAI-compatible
API, durable automation runs, local knowledge and learning, source-improvement
candidates, and optional adapter training. Your node runs on your own Linux
machine or server. There is no required TalkToAI server, central registration,
main node, or company-hosted database.

Source ·
Releases ·
Install guide ·
Workbench ·
Node independence

What changes in 7.3

Fresh installations use a local endpoint, an empty remote Hive endpoint,
local mode, and sharing disabled.
Upgrades remove exact retired owner-service defaults, preserving explicit
user-owned endpoints and existing private state.
Installers, checksums and runtime ZIPs come from versioned GitHub release
assets. Mirrors can be operator-selected; GitHub is a distribution host,
not an inference or federation coordinator.
Tab Pilot 0.3.1 is distributed as a verified GitHub ZIP. Linux and Windows
helpers prepare it for explicit Load unpacked approval, without forcing
a browser policy or depending on an owner-hosted extension update service.
The 7.2 improvement workspace remains: editable candidates, diffs, recorded
checks, digest-bound apply and rollback; small adapter fixtures and optional
local PEFT training do not automatically replace inference weights.

Run your own AI machine

Capability	Local implementation
Chat and models	Ollama routing, Ministral 8B runtime-template default, Gemma fallback, CPU profiles
Coding and tools	Files, bounded commands, archives, web reading, configured search, SSH tools
Durable runs	Checkpoints, budgets, stop/revoke, restart recovery, exact-action confirmations
API	Local OpenAI-compatible `/v1/chat/completions` with hashed machine keys
Browser control	Optional per-tab Tab Pilot grants and visible approval/stop controls
Knowledge	Local learning and operator-selected sharing; no mandatory remote catalogue
Source improvement	Separate source candidates, checks, reviewed apply and rollback
Training	Synthetic CPU fixture and optional LoRA/PEFT jobs with held-out evaluation
Voice	Optional local Piper/Voicebox; optional transcription dependencies
Offline operation	Bundle code, dependencies and models before disconnecting

Models and software dependencies require downloads during online installation.
Optional web search, cloud APIs, remote models and operator-selected federation
need their respective external services. Local inference is not distributed
training, peer-to-peer GPU pooling or a guarantee of AGI.

Install with checksum verification

On a Linux machine you administer, review the script before running:

mkdir -p openzero-bootstrap && cd openzero-bootstrap
curl -fL https://github.com/ResearchForumOnline/OpenZero/releases/download/v7.3.0/install.sh -o install.sh
curl -fL https://github.com/ResearchForumOnline/OpenZero/releases/download/v7.3.0/install.sh.sha256 -o install.sh.sha256
sha256sum -c install.sh.sha256
less install.sh
bash install.sh --desktop

For a headless node use --server --no-tab-pilot; --skip-model avoids the
large default-model download. The runtime archive is separately verified before
unpacking. A checksum detects corrupt/mismatched assets; it is not a detached
publisher signature. The default runtime and extension assets are pinned to
v7.3.0 so newer release assets cannot silently change this installer.

Open the panel at http://localhost:1024. Keep it on loopback, or access your
own remote node through an SSH tunnel/VPN. Never put private tokens into GitHub.

For a pre-7.3 installation, use the verified installer above with --dir set
to its existing path. This replaces the updater that previously depended on
the retired owner host. Once updated, use the local updater:

cd ~/openzero
bash update.sh --server --no-tab-pilot

The updater retrieves the latest installer and its checksum from GitHub.
OPENZERO_INSTALLER_BASE_URL selects your own installer mirror;
OPENZERO_RELEASE_BASE_URL selects your own runtime/extension asset mirror.
The mirror must provide the same filenames and matching checksum files.
For source development, clone this repository, create a Python virtual
environment, install openzero/requirements.txt, and run openzero/run_brain.sh.

Browser extension

Tab Pilot source and setup ·
Verified ZIP ·
ZIP checksum ·
Windows helper ·
Linux helper

The extension plans through your own loopback node; a remote node needs your
own SSH tunnel. Generate a scoped token in the local panel and enter it in
extension Options. Browser approval remains required. Existing managed installs
are not automatically uninstalled; their old policy should be removed by their
administrator before moving to the unpacked package.

Models and research

The default is
OpenZero Ministral3 8B Runtime Agent
through Ollama as
hf.co/shafire/OpenZero-Ministral3-8B-Runtime-Agent-GGUF:Q5_K_M.
It uses unchanged upstream weights with an OpenZero runtime template.
Gemma — Compatibility fallback,
model guide,
research archive, and
Agentic GGUF collection
remain separate public resources.

Documentation and preservation

ZeroThink and Matrix bridge code is optional integration code for an
operator-owned deployment, not a requirement to use a retired public website.
Old release notes describe historical services; current installation guidance
is this README and the 7.3 guides.

License and privacy

Source is available under the OpenZero Community Source License.
Inspection, personal use, internal business use and contributions are permitted
under its terms; commercial resale or managed hosting needs written permission.
This is not an unrestricted open-source licence.

Release archives exclude .env, credentials, generated keys, databases,
private uploads, model stores, runtime knowledge and backups. Do not expose the
operator panel publicly without strong authentication and network controls.
See SECURITY.md.