[Go to site: main page, start]

AI for defense, inside your boundary.

Give operators one governed place to use AI without creating a new cloud dependency, vendor-controlled data path, or cross-enclave workflow. Open WebUI runs inside your boundary, connects to the models you approve, and leaves data movement to your explicit routing decisions.

Air-gap ready
Run fully disconnected when configured with offline mode, local models, and no external providers.
Classification boundaries
Deploy one instance per enclave so data, models, and logs remain inside the boundary you operate.
Identity & access
Map SSO, groups, and role-based access control to how your organization actually works.
Audit & accreditation
Keep logs, retention, and administrator review inside your own accreditation process.

“Readiness is not connectivity. It is knowing your tools still answer when every link is cut, your data sits where you left it, and nothing reaches the mission without your order. The network can fall. The mission cannot.

The network
370M+
downloads
471K+
community members
148K+
GitHub stars

Defense-grade is control across the whole workflow.

01
Enclave

Run on the infrastructure you have already accredited: on-premises, in a sovereign cloud, at the tactical edge, or fully air-gapped. Docker, Kubernetes, or the process your environment already trusts.

02
Data path

Prompts, files, retrieval results, embeddings, and logs can remain inside the boundary when you connect only approved local services. Data crosses your network edge only when you explicitly route it there.

03
Command & audit

Identity, access, retention, and oversight can stay with your people. Every line of source is open to inspection, so reviewers are not limited to a black-box workspace.

The hard part is not only where the model runs.

Operators need

A usable place to chat, work with documents, search knowledge, call tools, and share results.

Security needs

Accreditation, access mapped to role and need-to-know, retention, audit trails, and administrator review.

Split across unmanaged tools, that gap becomes shadow AI. On a classified network, shadow AI is not a productivity problem. It is an incident.

Open WebUI gives the mission one governed place to use AI.

Deploy it inside the wire. Connect the models your environment approves. Keep chat, files, permissions, and administration in one self-hosted workspace your security team can actually see.

Deploy disconnected

Run air-gapped with local engines like Ollama or vLLM, with external providers disabled.

Choose models

Use any approved model, open-weight or hosted, and swap as requirements change.

Inspect everything

Review the source, deployment pattern, scanner guidance, and vulnerability process before approval.

Forward-deployed teams need AI that keeps working when reachback is denied, disrupted, intermittent, or limited. Local compute, local models, local control.

Security review materials for the evaluation file.

The controls that matter in a defense environment should be inspectable before a pilot or ATO discussion begins. Use these materials to brief deployment boundaries, data handling, identity, audit, hardening, and supply-chain review.

Get started

Deploy it inside the wire.

$ pip install open-webui

One command. 60 seconds. No account required. Run it with Docker, Kubernetes, or your existing deployment process.

Contact enterprise sales

Defense AI, answered.

Can Open WebUI run on classified or air-gapped networks?
Open WebUI is built to be self-hosted and can run inside a disconnected environment when configured with offline mode, local models, and no external providers. Paired with a local inference engine like Ollama or vLLM, prompts, files, retrieval, and model calls can remain inside the network you operate. The standard pattern is one deployment per enclave, so classification boundaries are enforced by infrastructure, not policy. Whether it operates on a given classified network is your accrediting authority’s decision; the software adds no required external dependency that stands in the way.
Is open source safe for defense use?
Open source is one part of the assurance story: it makes the running code inspectable. The defense posture still depends on disciplined deployment: pin versions, review dependencies, scan artifacts, restrict privileged capabilities, and operate the stack inside an accredited boundary. Open WebUI supports that review model because the source, deployment path, security policy, and scanner guidance are public rather than hidden behind a black box.
Does anything leave the enclave?
In an offline deployment configured with local models and no external providers, chats, files, knowledge bases, embeddings, users, permissions, and logs sit in the database and storage you point Open WebUI at, inside your own infrastructure. There is no usage telemetry, and offline mode disables outbound conveniences like update checks and model downloads. Data crosses your boundary only if you explicitly route it there, such as connecting a hosted model provider.
How does Open WebUI fit our accreditation process?
Open WebUI deploys into the environment you have already accredited and inherits its controls: your hardened hosts, your network segmentation, your identity provider, your logging pipeline. In an offline local-model configuration, it does not require an external service, outbound model data flow, or runtime vendor dependency inside your boundary. Access maps to your SSO and role-based access control, and audit, retention, and review can be aligned with your existing process.
Which models can we use in disconnected environments?
Any model you can serve inside the boundary. Open WebUI connects to local open-weight models through Ollama, vLLM, or any OpenAI-compatible endpoint, in any combination. As the open-weight landscape changes, you swap models without changing the workspace, so the capability your teams rely on is never tied to a single vendor, provider, or jurisdiction.
Can Open WebUI operate at the tactical edge or in DDIL conditions?
Yes. Open WebUI runs on anything from a single laptop to a Kubernetes cluster, which is what makes it viable in denied, disrupted, intermittent, and limited-bandwidth environments. A forward-deployed instance with a local model needs no reachback: it keeps answering when the link goes down, and everything it produces stays on hardware you carry.