Launching today
Lukan AI Agent, IDE and workstation.
The open-source AI workstation for coding, ops, and life
36 followers
The open-source AI workstation for coding, ops, and life
36 followers
Lukan is an open-source AI workstation that runs on top of your operating system to complete real end-to-end work. It includes an integrated coding agent that can read and edit files, run terminal commands, and automate technical tasks. Its terminals are tmux-based for persistent sessions, and you can run multiple agents either directly in terminal panes or through the agent UI. You can also create agentic pipelines to automate multi-step workflows with approvals and notifications.
Lukan AI Agent, IDE and workstation.
the pipeline approvals layer - is it blocking or async? curious how you handle failure recovery when an agent step needs human input but the user is offline
Lukan AI Agent, IDE and workstation.
@mykola_kondratiuk It's blocking per-agent but async globally. Each agent loop awaits its own approval channel when a tool needs approval that specific turn blocks until a response comes in (or cancellation). But the daemon runs multiple agents in parallel, each with its own channel, so one waiting agent doesn't block others. When the user is offline pending approvals are queued per-session in the daemon. When the client reconnects, we resend the approval required events so the user sees what's waiting, once approved, the channel unblocks and the turn resumes. Some issues whe daemon restart, in-memory state is lost, but you can re-issues the request. Pipelines runs are persisted to disk as json.
solid pattern. blocking per-agent keeps each loop predictable without stalling the whole system. the approval channel approach is what separates it from trust-the-model setups. most multi-agent frameworks skip this entirely.
Lukan AI Agent, IDE and workstation.
@mykola_kondratiuk Exactly, and that's also why in pipelines you can configure which tools each step has access to. Instead of giving the agent the full toolbelt on every step, you scope it down to just what that step needs: one step might only need ReadFiles + Grep, another just Bash with specific patterns, another only WebFetch. It's least-privilege by design rather than trusting the model to "do the right thing" with everything enabled.
per-step tool access is underappreciated. most frameworks give the whole pipeline the same permissions. thats where things get sloppy fast.
ProdShort
Nice, It feels like OpenClaw on steroide. Who is your target ?
I feel that every tool online tried to make Openclaw easier, I feel you are trying to make it more powerfull instead.
Lukan AI Agent, IDE and workstation.
@bengeekly Yes, the idea is to enhance it, if you need skills or plugins, you choose; if you need sandbox isolation, you activate it; if you need internet access, you can use Lukan Relay (a Rust binary that has Google login and device login in the machine implemented and connects to the daemon via WebSocket), tmux-based terminals to persist them, etc.