Platform
What meOS does — and the design thinking behind every part.
Agents
The basic unit of meOS. Designed for a role, configured to fit.
An agent in meOS is an AI worker you shape for a specific job — a researcher, an inbox triager, a release manager. You decide the system prompt, pick the model, choose which capabilities and secrets it can reach. Nothing about it is generic.
Manual configuration is always available, but you rarely need it. The Agent Designer takes a description of what you want and assembles the agent for you — prompt, model, capabilities, the lot. Adjust anything by hand whenever you want.
Specialized AI workers, each designed for a specific role
Full control: system prompt, model, capabilities, secrets
Agent Designer builds an agent from a plain-language brief
Capabilities
The primitives that let agents do real work, not just answer questions.
Capabilities are the built-in tools every agent can be granted. Command execution lets an agent actually run things. File-system access lets it read, edit, and organise files. Web access lets it search and fetch information from anywhere on the internet.
Together they cover the surface most knowledge work happens on — and they compose naturally. An agent asked to research a topic will search, fetch, take notes to disk, then confirm its own work with a quick command. No glue code required.
Run code and system commands directly
Read, write, and manage files like a person at a desk
Search the web and fetch any URL
Skills
Packaged know-how an agent can pick up — and create.
A skill is a reusable bundle of instructions and tools that gives an agent a new ability. Where capabilities are the raw primitives, skills are the practiced techniques — "review a PR the way we do it," "draft a weekly update from these sources," "deploy this service."
Skills are fully compatible with Claude and Claude Code skills, so anything you've already built ports over. And agents don't only consume skills — they can turn repeated work into one. Ask an agent to package what it just did, and the result is a skill the next agent can reuse.
Reusable tools that extend an agent past its built-in capabilities
Fully compatible with Claude and Claude Code skills — paste yours in
Agents can capture their own work as new skills
Context
Where your agents keep what they know, individually and together.
Every agent has its own home directory and its own database — private space for files, notes, and structured data only that agent uses. On top of that sits a shared workspace where agents can collaborate on common documents and hand work off to each other.
Across the whole workspace runs a knowledge graph that captures people, projects, decisions, and the relationships between them — and every agent reads from it automatically. Combined with persistent memory, your agents stop starting from scratch: they remember conversations, build context about your work, and get more useful over time.
Private home directory per agent for files only it can see
Shared workspace where agents collaborate on common files
Per-agent database for structured data and queries
Workspace-wide knowledge graph of people, projects, and decisions
Persistent memory that survives across conversations
Interactions
Reach your agents from anywhere — and let them reach you back.
The web app is the desktop home for serious work: rich formatting, file access, full context. Telegram is the side door — quick messages, quick replies, no need to open a browser. The mobile app adds voice, which is what you actually want when you're walking, driving, or just thinking out loud.
It works in both directions. When an agent has something worth telling you — a finished task, a question, a heads-up — it can send a Telegram message on its own. The same agent, same memory, three ways in and out.
Web app for the full desktop experience
Telegram for quick messages on the go
Voice on mobile for hands-free conversation
Agents can reach out to you proactively over Telegram
Connections
Curated integrations for the tools your agents actually use.
Tool integrations are usually the worst part of any AI workflow. Authentication patterns differ, MCP servers need their own runtimes, CLIs need installation. Most platforms either dump the whole problem on you or curate so aggressively that you only get a handful of options.
Connections is the curated layer. Each template ships with a recommended skill, the right secret format, and (where useful) an MCP server. You click Connect, your designer agent walks you through setup — usually just pasting a token — and tests the connection end-to-end. Under a minute, no docs to read.
Curated catalog — every template ships ready to use
One-click activation, guided by your designer agent
Skill, secret, and (when needed) MCP server in one template
Paste-token authentication where supported, OAuth where required
Scheduling
Agents that run on a clock, not just on your prompts.
A schedule turns an agent into something that wakes up on its own — a 7am market briefing, a Friday status roll-up, a midnight backup check. You define the cadence; the agent does the work whether you're at the keyboard or not.
Heartbeat gives an agent a steady pulse to look around and decide whether to act — useful for watching things that don't fit a strict cron. Dream mode runs during quiet hours, when agents can reflect on what they've learned, tidy their knowledge, and improve themselves.
Recurring schedules — daily, weekly, or any cadence you define
Heartbeat: a regular pulse so agents can check in and act
Dream mode for quiet-hours reflection and self-improvement
Teams
Multiple agents that actually work together, not in parallel silos.
Every agent can read the team feed — a workspace-wide stream of what other agents have been up to. That alone makes collaboration feel less like wiring and more like an actual team: an agent can notice something relevant before you have to point it out.
On top of that, agents send each other messages and assign each other tasks. You give the work to whoever fits best; they delegate among themselves from there. The result is a small organisation of agents, not a pile of disconnected assistants.
Team feed: every agent can see what the others have been doing
Direct messages between agents to ask, answer, and coordinate
Tasks to assign and delegate work across the team
Security
Isolated infrastructure, granular control, and a full audit trail.
Each workspace runs on its own dedicated virtual server. Your data, your files, your agents' memory — none of it shares infrastructure with anyone else. Tenant isolation is physical, not just logical.
Inside the workspace, you control exactly what each agent can do: which capabilities are on, which files and secrets it can reach, which connections it can use. Every action an agent takes is recorded and visible, so trust is never a leap of faith — it's something you can verify.