FAQ
Questions that come up repeatedly. Longer answers live on the page that owns the topic; these are the short forms.
Do I need a GPU?
Only for local inference or local embeddings. Arble itself is a CPU workload — point it at a hosted provider and you need no GPU at all.
Does Arble send my data anywhere?
Only where you configure it to. A local model means prompts never leave the machine. A hosted model means prompts and completions go to that provider. Memory, audit logs and session history stay local either way.
Can I use it fully offline?
Yes, with a local model. Connectors and remote MCP servers need network access; sessions, memory, permissions and desktop control do not.
How do I stop an agent mid-run?
Ctrl-C in the foreground, or arble agents kill <id> for a background one. Both
cancel cooperatively — an in-flight tool call gets the chance to stop cleanly.
Why did my run stop early?
Usually a denied permission rather than a bug. Check
arble permissions audit before reading the trace.
Does deleting a session delete its memory?
No. Memory is project-scoped and outlives sessions deliberately. Delete memory entries explicitly.
What is the difference between a tool and a skill?
A tool is one capability. A skill is a procedure that calls tools. Skills describe how your team works; tools describe what is possible.
Do I need MCP if I use the Tool SDK?
No, and they solve different problems. MCP connects Arble to a server someone else runs; the Tool SDK builds a tool that runs as part of Arble. Many setups use both.
Can two people share a project?
Yes. Projects belong to the organization; memory and permission grants are shared. Sessions are per-user unless explicitly shared.
Is the CLI output safe to parse?
Parse --json, not the rendered tables. Table layout changes between releases; the JSON
shape is versioned.