A morning research digest
New papers read overnight, summarized by morning.
Private AI agents. Encrypted by default.
The agent OS for home and office,
The agent OS for home and office,
on hardware you own.
Agents run as durable services on your own machine: they keep working while you sleep, put a face in the room, and a network capture shows nothing leaving.
Self-hosted AI
Local by default
Encrypted chats
Vault Password
Incognito chats
Egress Audit
Encrypted memory
No telemetry
Encrypted files
Local LLMs
3D avatar
Agents are the processes. Scheduled tasks are the services. Skills and tools are the drivers. The avatar is the display server.
Why Oscova Agents
Built to run without the cloud, and to prove it.
Runs network-denied. Still works.
Block the network and agents keep working on local models. The Egress Audit lists every moment data ever left, and the Session Inspector shows a run stayed local.
Speech in and out, offline.
Whisper listens and Kokoro speaks, both running offline. Cloud speech is an opt-in provider, never the default.
It works while you sleep.
Agent tasks run overnight: research, a morning digest, a watcher. The work costs electricity, not metered tokens.
How it works
From install to assistant in minutes
Install once, pick a model, start talking, or hand the work to scheduled agents.
Install in one command.
The oscova global tool installs the app, the agents, and the Oscova
Avatar Engine together.
Pick your model.
Local by default, cloud by choice. Pin it so today's model is tomorrow's model.
Talk now, or schedule it.
Chat in the moment, or schedule agent tasks overnight and wake up to results.
What's inside
Built around local models and provable privacy.
What the runtime does out of the box, from local language models and encrypted chats to the face it puts on screen.
Local models, first-class
Run local large language models on your own hardware, built lean so small models do real work. Cloud stays opt-in.
Encrypted by default.
One toggle keeps runs offline. Every agent's chats, tool calls and task runs are encrypted under your Vault Password. Definitions, user settings, memory and search indexes are encrypted too.
The whole household, one box
Each person gets their own identity, tasks, and conversations. Any browser on your network can connect.
A face, not a tab
The Oscova Avatar Engine streams an expressive presence to any browser on your network, subtitles included.
Extend it your way
Add capabilities with MCP tool servers, skills, and NuGet plugins an admin approves first.
Bring your own channels
Bind agents to Discord or Telegram bots and reach the same assistant from anywhere.
Your data, as files you own
Chats are plain files in your own folder, not rows in a database. One passphrase-protected backup moves the whole environment to a new machine.
Agents, run as services
Agents carry their own instructions, tasks, and run history you can inspect. Services, not prompts in a chat box.
Routines you can set up.
Not customer stories: agent tasks you can set up yourself, running on your own hardware while you get on with your day.
An inbox, triaged before you sit down
Overnight mail read and sorted, a short brief waiting in your notifications.
A confidential case file, kept in cipher
Notes and documents summarized by an agent into encrypted files you unlock with your Vault Password.
A daily plan, ready
Each person's day, drafted before they wake.
Private AI research
Research that stays in the building.
Point an agent at your notes, papers, and files. It reads, summarizes, and drafts on your machine, with a local model if you want one. Nothing is uploaded to us, because there is no us in the middle.
-
Every agent's chats, tool calls, task runs and memory are encrypted under your Vault Password.
-
Incognito for the question you do not want kept at all. Never written, never remembered.
-
The Egress Audit lists every call that left the machine. Private means checked, not promised.
Go deeper
Everything the runtime does is documented.
From first install to writing your own plugins: the user guide and developer guide cover every screen, concept, and extension point.
Agents
Create agents with their own instructions, tools, skills, memory, and access, then share them.
Tools and skills
Tools are the actions agents can call. Skills are packages of instructions they load when needed.
Plugins
Extend the runtime with NuGet plugin packages that run only after an admin enables them.
Connections
Bind agents to Discord and Telegram, add MCP tool servers, and connect Google or Microsoft accounts.
Governance and admin
Accounts, roles, and sign-in for the household, plus control over which tools each agent may use.
Developer guide
Build your own toolsets, skills, and plugins in C# on the same runtime the app uses.
This is not our first assistant. It is roughly our seventh, twenty years in the making. Read the story
FAQ
The essentials, before you install.
Straight answers on proof, confidentiality, hardware, cost, and how it was built. Everything else is in the docs.
By default, nothing. Block the network and it keeps working on local models. When you do allow a cloud model or an Internet tool, the Egress Audit records each call with the agent, provider, and tool involved, and the Session Inspector shows the run. The proof is behavioral, not a checkbox.
That constraint shaped the architecture. The system prompt stays lean, tools and skills are discovered just-in-time instead of dumped into context, and long sessions compact automatically. Small models do real work on hardware you may already own; cloud providers remain an opt-in choice.
Conversations are encrypted files under your own user folder, one transcript per chat. You can back them up or delete them with a file manager. Oscova opens their content after you unlock your Vault Password. An administrator can export the environment as a passphrase-protected backup. Start a chat as Incognito and it is never written to disk at all.
It was designed for that. Nothing leaves the machine unless you allow it, conversations are encrypted at rest under your own key, there is no vendor account, and there is no telemetry. A clinic, a legal practice, a finance team, or a research lab under NDA gets the same runtime as a household. We describe the behavior, not a certification, so check it against your own policy.
The runtime and the two built-in avatars are free. No subscription, nothing to buy to get started. Premium avatar options come later.
Yes. Give an agent your papers, notes, and files and it reads, summarizes, and drafts on your machine. Every agent's chats, tool calls, task runs and memory are encrypted under your Vault Password. Pick a local model and nothing leaves at all. Private here means you can check it: the Egress Audit lists every call that went out.
No. Oscova Agents is built in C#, a language we know deeply. Human judgment guides the work. We read the code, refine it and review it ourselves. We have been building assistants for close to two decades, and this is roughly the seventh.
Install Oscova Agents with one command.
PowerShell · Windows x64
irm https://oscova.ai/install.ps1 | iexTerminal · macOS Intel or Apple Silicon
curl -fsSL https://oscova.ai/install.sh | bashTerminal · Linux x64/glibc
curl -fsSL https://oscova.ai/install.sh | bashAutomatic .NET setup currently supports Ubuntu 24.04.
If .NET 10 is needed, the installer will offer to install it first.
Once installed, run oscova start.
Let your coding agent install it. Copy the prompt and paste it into an agent that can run commands.