Machine
The Linux computer your agent runs on. It is created for you when you sign up, it can be started and stopped, and what it writes to disk is still there next time — in the next turn, in the next chat, tomorrow. The agent’s home is/home/daytona; the root filesystem is read-only. One machine runs one harness
at one ref, so a second harness of yours needs a machine of its own. How many
you may have, and whether they stay up when you are not talking to them, comes
from the plan. → Machine
Harness
The agent’s own git repository: the code that decides who the agent is. Programs, agents, hooks, tools, evals and the documentation the tree carries, with a version history where every version is published and immutable. The checkout lives at~/harness on the machine and is edited with ordinary file tools;
git push origin main is not a push but a publish. → Harness
Program
A directory underprograms/ in the harness with a main.py that declares the
agents in code. programs/splox/ is the one a chat is answered by, and its
handle(msg) says which agent takes the message. A program can also be a
long-running loop — a Telegram bot, a watcher, a nightly job — started once and
still going in tomorrow’s chat. → Program
Agent
Declared inside a program:agent("Assistant", system_prompt=…, model=…, tools=[…]). The platform asks main.py who an agent is at the moment it needs
to know rather than storing a copy, so an edit to the file lands on the next
turn. An agent listed in another agent’s agents=[…] becomes a tool of that
agent. → Agent
Run
One execution of an agent: it has a status, events, messages, outputs, a tree of sub-agent runs and usage. Every turn of a chat is a run, every sub-agent is a run, every call from a program is a run — they are all billed, listed and recorded the same way. → RunChat
A conversation in the app. It belongs to one machine and stays on it; a turn of it is a run. A chat can be renamed, forked from any message, moved into a project, shared as a public URL, or deleted. → ChatHook
A Python file atprograms/<name>/hooks/<point>.py that answers one point of the
run loop — context, model, tools, guard, memory, errors, stop and the rest. Hooks
are how a run is shaped without touching the agent, and they belong to the
program whose directory they sit in. → Hooks
Tool
A Python function the agent imports and calls in code, rather than a JSON call it emits. Platform packages arrive projected attools/<service>/<tool>.py, a
connected MCP server appears the same way, and a file you write at
tools/<name>.py in the harness becomes tools named <file>__<function>. What a
turn may call is decided by the agent’s tools=[…], the account’s system-tools
switch and hooks/tools.py. → Tools · Tool catalog
Skill
A folder with aSKILL.md that the agent reads when the task calls for it. The
description in its frontmatter is the whole selection mechanism — it is what the
model sees when deciding whether this skill is relevant — and the body is the
instructions it follows once it opens the file. Skills come from the library or
from your own harness. → Skills
Starter
A working harness you copy to get your own. Copying one creates a repository at its current commit and a machine to run it on; from that moment it is yours and diverges however you edit it. The provenance is recorded on the first version, so you can always merge later changes from the starter you came from. → StartersVersion
A published state of a harness, numbered, atrefs/versions/N in the repository.
Publishing is a push to main, which is checked before it becomes a version:
files in places a harness has no room for, a tree with no program, tools that
would not import, two tool files that would be shown under one name, an eval case
naming a grader that is not there. A machine runs main unless you pin it to a
ref. → Publishing a version
Window
One of the two rolling periods a plan grants — five hours and a week. Both measure what your runs cost, both are read as a percentage of the plan’s cap, and a full one is refused with the moment it frees up. The 5-hour window bounds a burst; the weekly one bounds living at that burst indefinitely. Windows belong to the account, so a new chat or a second machine does not reset one. → Usage · When a limit is reachedEndpoint
Where a model call is sent, and which credential pays for it. A platform endpoint is Splox’s own account with a provider; an endpoint you add on Connections is your key and your account. An agent names a model plus exactly one of a provider slug or an endpoint id, and Harness decides in the composer means no selection at all — the harness picks. → Model · ConnectionsInteraction
A run stopping to ask a person something. The agent raises one, the run’s status goes towaiting, and nothing happens until somebody answers — in the app, or
through the API. A confirm surface becomes a choice; a form becomes text. →
Interactions
Workspace
The git checkout an agent works in. The caller’s is/home/daytona/workspace; a
sub-agent spawned with workspace="landing-ui" gets a git worktree of its own at
/home/daytona/workspace-landing-ui, on a branch of that name, so no two agents
ever write the same tree. The files are at that path the moment the sub-agent
finishes — there is nothing to fetch and nothing to serialize. →
Sub-agents
How it fits together
The same words, in the order they depend on each other.
What changed
If you are here from the old documentation, start with this.