A secure, self-hosted personal AI agent. Handles calendar, email, files, web research, Telegram, and more - controlled by you, running on your own hardware.
| `image.gitlab.pm/rune/oai-web:latest-no-browser_arm64` | arm64 | No |
**Full image** includes Playwright and a Chromium installation. This adds roughly 350 MB to the image size but enables the browser tool - the agent can fetch pages, take screenshots, click elements, fill forms, and navigate web UIs on your behalf.
**No-browser image** is leaner and faster to pull. All other tools and features are identical. Choose this if you don't need the agent to interact with web pages directly, or if you are constrained on disk space.
# Access key for the Brain MCP endpoint (generate with: openssl rand -hex 32)
BRAIN_MCP_KEY=
```
Change the `change-me` passwords in `AIDE_DB_URL` and `BRAIN_DB_PASSWORD` to something strong. They must match - if you change `BRAIN_DB_PASSWORD`, the same value is substituted into `BRAIN_DB_URL` automatically.
---
## Personalising the Agent
### SOUL.md - Agent identity and personality
`SOUL.md` defines who your agent is. The name is extracted automatically from the first line matching `You are **Name**`.
Key sections to edit:
**Name** - change `Jarvis` to whatever you want your agent to be called:
```markdown
You are **Jarvis**, a personal AI assistant...
```
**Character** - describe how you want the agent to behave. Be specific. Examples:
- "You are concise and avoid unnecessary commentary."
- "You are proactive - if you notice something relevant while completing a task, mention it briefly."
- "You never use bullet points unless explicitly asked."
**Values** - define what the agent should prioritise:
- Privacy, minimal footprint, and transparency are good defaults.
- Add domain-specific values if relevant (e.g. "always prefer open-source tools when suggesting options").
**Language** - specify language behaviour explicitly:
- "Always respond in the same language the user wrote in."
- "Default to Norwegian unless the message is in another language."
**Communication style** - tune the tone:
- Formal vs. casual, verbose vs. terse, proactive vs. reactive.
- You can ban specific phrases: "Never start a response with 'Certainly!' or 'Of course!'."
The file is mounted read-only into the container. Changes take effect on the next `docker compose restart`.
---
### USER.md - Context about you
`USER.md` gives the agent background knowledge about you. It is injected into every system prompt, so keep it factual and relevant - not a biography.
**Identity** - name, location, timezone. These help the agent interpret time references and address you correctly.
```markdown
## Identity
- **Name**: Jane
- **Location**: Oslo, Norway
- **Timezone**: Europe/Oslo
```
**Language preferences** - if you want to override SOUL.md language rules for your specific case:
```markdown
## Language
- Respond in the exact language the user's message is written in.
- Do not assume Norwegian because of my location.
```
**Professional context** - role and responsibilities the agent should be aware of:
```markdown
## Context and background
- Works as a software architect
- Primarily works with Python and Kubernetes
- Manages a small team of three developers
```
**People** - names and relationships. Helps the agent interpret messages like "send this to my manager":
```markdown
## People
- [Alice Smith] - Manager
- [Bob Jones] - Colleague, backend team
- [Sara Lee] - Partner
```
**Recurring tasks and routines** - anything time-sensitive the agent should know about:
```markdown
## Recurring tasks and routines
- Weekly team standup every Monday at 09:00
- Monthly report due on the last Friday of each month
```
**Hobbies and interests** - optional, but helps the agent contextualise requests:
```markdown
## Hobbies and Interests
- Photography
- Self-hosting and home lab
- Cycling in summer
```
The file is mounted read-only into the container. Changes take effect on the next `docker compose restart`.
Set your personal bot token under **Settings - Telegram** (or **Settings - Profile - Telegram Bot Token**) if you want your own Telegram bot. Your chat ID must be whitelisted by the admin before messages are processed.
- **Trigger account** - dispatches agents based on keyword rules in incoming emails
- **Handling account** - a dedicated AI agent reads and handles each incoming email
### Browser - Trusted Domains
When the agent uses the browser tool to interact with a page (clicking, filling forms), it will ask for confirmation unless the domain is pre-approved. Add trusted domains under **Settings - Browser** to skip the confirmation prompt for those sites.
oAI is free software licensed under the **GNU Affero General Public License v3.0 (AGPL-3.0)**.
This means you are free to use, study, modify, and distribute oAI, but any modified version you run as a network service must also be made available as free software under the same license.
See [LICENSE](LICENSE) for the full license text, or visit [gnu.org/licenses/agpl-3.0](https://www.gnu.org/licenses/agpl-3.0.html).
oAI-Web takes real actions on your behalf - it can send emails, write files, make calendar changes, post Telegram messages, and interact with web pages. Review your whitelist and permission settings carefully before use. Content you send is processed by your configured AI provider (Anthropic, OpenRouter, or OpenAI). oAI-Web is provided "as is" without warranty of any kind - the author accepts no responsibility for actions taken by the agent or any consequences thereof. See LICENSE for full terms.
