This document answers one question only: "what will I actually be
able to do, in practice, from day one?". No architecture, no
code. Ten minutes, and you know what to expect.
If you haven't read Architecture
— Introduction (in Italian for now; English coming soon), you can still
follow this document, but the terms sandbox, workspace,
autonomy are defined canonically there.
The five things you get on day 1
Chat from the terminal.mykleos chat
opens a local text session. It is the fastest way to talk to the agent,
and by default runs at Supervised autonomy (it asks confirmation
for serious actions).
Chat from Telegram. A dedicated bot, to which you
connect once through pairing. After that, writing to the bot is
equivalent to talking to your home computer from anywhere.
Run shell commands inside a sandbox. Examples:
check how much disk space is left, search for a file in the workspace,
recompile a project, answer "what is the NAS IP address?".
Read and write in the workspace
(/opt/myclaw/workspace/). Here the agent has free rein (always
sandboxed). Outside the workspace, approval is required.
Automate repeated chores with cron. Example: "every
evening at 22:00 summarise the systemd log and send me a Telegram message
if anything unusual shows up".
A typical dialogue
Scenario: you're at the supermarket, you want to know whether the
Claude model download you started last night has finished.
Roberto
is the model download done?
Mykleos
Checking ~/downloads/ and the wget log. Give me a second.
(internal)
Policy: "fs_read + web_fetch to local log" → Supervised OK, no
approval needed. Bwrap started, read-only profile.
Mykleos
Yes. 8.2 GB of 8.2 GB downloaded, SHA256 verified. The file is
claude-opus-4-7.gguf, finished at 03:14 tonight. Should
I move it to /opt/models/?
Roberto
yes
(internal)
Policy: moving out of the workspace → requires approval. The "yes"
just received counts as explicit approval.
Telegram bot token, external provider API keys, etc.
Chmod 600 required. Gateway restart.
Remember: Mykleos does not decide which model to run. The choice
lives in the implementation of the
LLM abstraction
(in the author's environment: suprastructure, config in
/opt/suprastructure/config/). A single place to decide "which
model runs in the house": Mykleos inherits that, like any other sibling
agent of the same environment.
Glossary
The complete glossary of the project — technical terms, components,
sibling projects, people and places — lives in a dedicated document:
→ Open the Glossaryabout 35 entries, alphabetical, with links to canonical sources.
A few entries that recur in the Survival Kit, at a glance:
The architectural pattern: Mykleos accesses language models (and STT/TTS/embedding) only through abstract interfaces, not through clients of specific providers. Makes the system provider-agnostic.
The implementation of the LLM abstraction present in the author's environment. In a different environment the role would be played by a different adapter.