roadmap
- 01
- 02
- 03
- 04
- 05
- 06
- 07
- 08
Run OpenClaw, the self-hosted AI assistant, from one Gateway on your own hardware to agents in your chat apps. Covers channels, workspace memory, automations, skills and plugins, and security.
107 of 107 questions in your plan: 0 known, 0 in progress, 0 due, 0 skipped, 107 new. 0 cut, outside your plan.
Interviewers reach for OpenClaw to test whether a candidate can run an agent that real people message every day, not only call a model from a script. Teams building internal assistants ask how its parts fit together: why one long-running Gateway owns every chat connection, what happens when a second message lands while the agent is still busy with the first, and why two people writing to the same assistant can share one conversation until `session.dmScope` separates them. Platform and infrastructure interviewers move to hosting: what keeps the Gateway alive after a reboot or an SSH logout, and how you reach its control UI on a rented server without opening the port to the internet. An assistant with shell access that reads group chats is a prompt-injection target by design, which is why security interviews dwell here longest: they want to hear how pairing and allowlists decide who can reach it, what `openclaw security audit` reports, and which of tool policy, the sandbox and exec approvals actually stops a hostile message from running a command. The agent loop, memory and context design underneath belong to the [AI Agents roadmap](/roadmaps/ai-agents), and [Claude Code](/roadmaps/claude-code) is the comparison interviewers like to ask for: a coding agent in a terminal against an assistant that lives in chat apps.
Treat the plan as a question bank to work through, with the [OpenClaw topic page](/topics/ai-openclaw) open beside it as the map of how the parts connect. Go back to that map whenever an answer names a config key but not the process that reads it. The level decides where to dig. For a junior, the questions that matter most sit in the [Gateway control plane](/topics/ai-openclaw-gateway-control-plane): what the single process owns, why it binds to loopback, and how a client or a paired device connects to it, the picture almost every later answer assumes. From middle level the weight shifts to [DM and group access](/topics/ai-openclaw-channels-access), pairing codes against allowlists and mention rules in a busy group, because that is where a careless setup starts answering strangers. The senior and principal questions gather in [sandboxing and tool policy](/topics/ai-openclaw-security-sandbox), where you weigh what a sandbox mode still leaves on the host against what a deny list removes outright. Several OpenClaw defaults have moved across recent releases, such as where heartbeat instructions live and whether direct messages share one session, and the answers name the release they assume. Before quoting a default in an interview, check it against the version the team actually runs.
Interviewers reach for OpenClaw to test whether a candidate can run an agent that real people message every day, not only call a model from a script. Teams building internal assistants ask how its parts fit together: why one long-running Gateway owns every chat connection, what happens when a second message lands while the agent is still busy with the first, and why two people writing to the same assistant can share one conversation until session.dmScope separates them. Platform and infrastructure interviewers move to hosting: what keeps the Gateway alive after a reboot or an SSH logout, and how you reach its control UI on a rented server without opening the port to the internet.
An assistant with shell access that reads group chats is a prompt-injection target by design, which is why security interviews dwell here longest: they want to hear how pairing and allowlists decide who can reach it, what openclaw security audit reports, and which of tool policy, the sandbox and exec approvals actually stops a hostile message from running a command. The agent loop, memory and context design underneath belong to the AI Agents roadmap, and Claude Code is the comparison interviewers like to ask for: a coding agent in a terminal against an assistant that lives in chat apps.
Treat the plan as a question bank to work through, with the OpenClaw topic page open beside it as the map of how the parts connect. Go back to that map whenever an answer names a config key but not the process that reads it.
The level decides where to dig. For a junior, the questions that matter most sit in the Gateway control plane: what the single process owns, why it binds to loopback, and how a client or a paired device connects to it, the picture almost every later answer assumes. From middle level the weight shifts to DM and group access, pairing codes against allowlists and mention rules in a busy group, because that is where a careless setup starts answering strangers. The senior and principal questions gather in sandboxing and tool policy, where you weigh what a sandbox mode still leaves on the host against what a deny list removes outright.
Several OpenClaw defaults have moved across recent releases, such as where heartbeat instructions live and whether direct messages share one session, and the answers name the release they assume. Before quoting a default in an interview, check it against the version the team actually runs.