DEMIROS
DemirOS is the secure, offline-capable OS that runs every Demirari AI machine. Deploy models, manage users, monitor hardware, and serve OpenAI-compatible endpoints — all from one dark, quiet pane of glass.
THE AGENT CONSOLE
Your private AI team, in one console.
Every Demirari system ships with the Agent Console — a self-hosted, OpenAI-compatible gateway that turns your local models into a working agent. Ask a question, attach a screenshot, or hand it a task: the console picks the right model, calls from 46 local-first tools, and streams back a sourced answer. No cloud provider, no third-party keys, nothing leaves the machine.
Research with receipts
Private metasearch and page fetch — every answer arrives with its sources, and no query ever leaves your network.
Real troubleshooting
A multi-step agent loop works through errors, configs, and logs the way an engineer would — cause first, then the fix.
Local image generation
Prompt to picture on your own GPU. No cloud image API, no uploads, no per-image metering.
Governed by design
Roles with per-role tool permissions, usage budgets, human-in-the-loop approvals, and an append-only audit trail.
IT troubleshooting
The agent walks through a real Power Platform error — likely cause, exact fix, and the known tool bug — step by step.
TOOL ARSENAL
46 tools on board. Zero in the cloud.
The agent decides for itself which tools a task needs, calls them in sequence, and shows its work. Every tool runs against your infrastructure or your accounts — and any Model Context Protocol server you mount appears as a native tool.
Research & web
Private metasearch, page fetching, and a full headless browser that can navigate, click, and screenshot.
web_search · fetch_url · browser
Documents & data
Drafts Word, PDF, and HTML deliverables; reads PDF/Office files with OCR; builds spreadsheets and diagrams.
docgen · document · spreadsheet · diagram
Media & vision
Local image and video generation, screenshot understanding, and speech-to-text — no cloud media APIs.
image_gen · video_gen · vision · stt
Knowledge & memory
Searches your private document base, remembers across conversations, and reuses saved playbooks.
rag · memory · skills
DevOps & network
Sandboxed shell, declared-host SSH, read-only SQL, git, DNS/WHOIS, ping/traceroute, packet capture, monitoring.
shell · ssh · sql · git · net_probe · pcap
Business systems
Microsoft 365, Google Workspace, Jira/GitHub ticketing, and webhook or email notifications.
m365_* · google_* · ticketing · notify
AUTOMATION
It works while you sleep.
The Agent Console is not just a chat window. It is an always-on operator you schedule, trigger, and delegate to — with named presets for recurring roles like researcher, SRE, extractor, or documentation writer.
Scheduled agent jobs
Recurring runs on a cron schedule — nightly log sweeps, morning briefs, watchdog checks — each with its own execution history.
Webhook triggers
Any system that can send an HTTP POST can start an agent run. A ticket lands, a monitor fires, the agent is already on it.
Sub-agent delegation
Large tasks split into focused sub-agents that work independent sub-tasks and report back — parallel work, one answer.
Memory & projects
Long-term memory across conversations, plus project workspaces with shared standing instructions, files, and cross-chat context.
CAPABILITIES
Everything you need. Nothing you have to babysit.
One-pane admin
Models, endpoints, users, and hardware health in a single dashboard.
OpenAI-compatible API
Drop-in endpoints; point your apps at the box, not the cloud.
Model management
Load, version, and roll back your Modelsmith weights.
Observability
Latency, throughput, GPU/thermal, and per-endpoint usage.
Access control
RBAC, SSO/SAML, and full audit logs.
Offline updates
Signed, air-gap-friendly updates via secure media.
ADMIN CONSOLE
Every request. Every tool call. Every decision — visible.
STEP 01
Monitor
Live request counts, token throughput, per-model latency, and failures — one quiet panel, no clutter.
STEP 02
Approve
Sensitive actions — like a database write — pause for a human decision. Approved or denied, the agent keeps working.
STEP 03
Control
Users, roles, and API keys live in the same pane. A role never sees tools it was not granted.
AGENT GOVERNANCE
Powerful doesn't mean permissive.
Every capability the agent has is something you deliberately granted — and can meter, gate, or revoke. Control lives at the platform layer, not in a prompt the model could talk its way around.
Per-role tool allowlists
A role is only ever shown the tools it is granted — the model cannot attempt what it cannot see.
Budgets & rate limits
Per-role token and request quotas on a rolling window, plus per-credential concurrency caps.
Human approvals, fail-closed
Gated tools pause for a human decision; no decision in time means denied — never silently approved.
Deep observability
Per-request analytics, Prometheus metrics, and OpenTelemetry trace waterfalls for every agent run.
CONNECTS TO YOUR STACK
Microsoft 365Google WorkspaceJiraGitHubPrometheusGrafanaSNMPMCP servers
SECURITY
Secure by default. Auditable by design.
DemirOS assumes the network is hostile and the audit is coming. Every control below is on out of the box — no add-ons, no enterprise tier.
- Zero telemetry / zero egressNothing phones home. Ever.
- RBAC + SSO/SAMLYour identity provider, your roles, your rules.
- Immutable audit logsEvery action, attributable and tamper-proof.
- Signed updatesOnly Demirari AI-signed builds can install.
- TPM + secure bootVerified from power-on to first token.
- FIPS 140-3 crypto module optionsDeployments can be configured with FIPS 140-3 validated cryptographic modules; certificate details provided during engagement.
FOR DEVELOPERS
If it works with the API, it works with your box.
Point your existing tools at your machine’s endpoint. SDKs for Python/JS, plus docs and examples.
- POST/v1/chat/completions
- POST/v1/embeddings
- GET/v1/models
- GET/v1/health
DEMIROS