The dashboard problem nobody solved: your AI agents are invisible
Cross-border sellers in 2025 run more autonomous software than most SaaS companies did in 2015. A single Shopify operator might have a Claude Code session rewriting product descriptions, a Codex CLI job cleaning supplier CSVs, a Gemini CLI process translating listings for EU marketplaces, and a Paperclip agent answering buyer messages on TikTok Shop. The bottleneck is no longer compute or model quality — it is visibility. When four agents are running in four terminal tabs, you cannot tell which one is waiting on your approval, which one silently failed, and which one finished an hour ago. That is the exact gap Cubicle, built by Çağlar Utku GÜLER, tries to close — and it is more relevant to Amazon FBA brand owners than the maker probably realizes.
What Cubicle actually solves
Strip away the pixel-art office and the product is a read-only status layer for AI agent fleets. Each running agent gets a “desk.” A working agent sits and types with its current task floating above its head. An agent that needs human input raises a hand. A finished agent walks to the lounge. An errored agent turns the screen red. The maker’s own framing is blunt: he runs “a small company of AI agents on Paperclip and could not see who was doing what without reading logs.”
The architecture is deliberately minimal. It is zero-dependency — one Node file and one HTML page, launched via npx @caglarutkuguler/cubicle. It works with Paperclip, Claude Code, Codex, Gemini CLI, or “any tool that writes a small JSON file.” It runs as a standalone page, a TV in kiosk mode, or a Paperclip plugin with a dashboard widget. There is a live demo at caglarutkuguler.github.io/cubicle with no install required.
The design constraint that matters most for operators: it is read-only by design. As the maker states, “Cubicle only reads. It cannot approve, run or change anything in your agent system.” The single optional write is a Telegram reply posted as a comment. Everything else — approvals, terminal prompts, Paperclip issue answers — still happens in the original tool.
Why Amazon sellers should care more than Shopify ones
Shopify operators tend to work in one browser tab with a handful of apps. Amazon FBA brand owners, by contrast, already live inside a fragmented stack: Amazon Seller Central for listings and inventory, Helium 10 for research, Klaviyo for email, a 3PL portal for fulfillment, and now increasingly an agent or two for copywriting, PPC bid suggestions, or review triage. The number of “who needs me right now?” surfaces is already high. Adding four CLI agents without a status layer is how sellers end up missing a permission prompt for six hours while a listing rewrite sits half-finished.
How it differs from the existing options
The obvious comparison is observability tooling. LangSmith and Langfuse are the grown-up answers here — trace-level logging, token accounting, evaluation harnesses. They are excellent if you are shipping an LLM product and need to debug retrieval pipelines. They are also heavy, developer-oriented, and priced for teams. Cubicle is the opposite: a glanceable ambient display for a solo operator or a small team.
The second comparison is generic dashboards — Grafana, Datadog, or even a Notion board you update by hand. Grafana can absolutely render agent state if you pipe metrics into it, but nobody sets up a Grafana panel to answer “is the translation agent stuck?” The friction is the point. Cubicle’s install is one npx command and one hook file.
The third comparison is the agent tools themselves. Claude Code, Codex, and Gemini CLI each have their own terminal output. None of them aggregate across tools. If you run three vendors, you have three silos. Cubicle is vendor-agnostic precisely because it reads a JSON feed rather than integrating deeply with any one runtime.
Where the math breaks
The read-only constraint is philosophically clean and operationally annoying. If an agent raises a hand because it needs approval, Cubicle tells you — but you still have to switch to the terminal or the Paperclip issue page to answer. For a seller running five agents, that is five context switches per approval cycle, and Cubicle does not reduce them. It only makes them visible. The maker is explicit: “Cubicle does not approve, continue or stop a run.”
There is also the multi-machine problem. A seller running agents on a laptop, a home server, and a VPS will need to reconcile three JSON feeds. The source does not describe a hosted aggregation layer, so this is a “not disclosed” gap rather than a solved feature.
What cross-border sellers can borrow from Cubicle
Even if you never install it, the design choices are worth stealing for your own ops stack.
Ambient status beats pull-based reporting. Most seller dashboards are pull-based: you open a tab, you check a number. Cubicle is push-based and ambient — a TV in kiosk mode in the corner of the office, or a second monitor. The lesson for DTC operators: your 3PL exception queue, your TikTok Shop order-hold queue, and your Amazon account health alerts should all be ambient, not something you remember to check.
Read-only is a feature, not a limitation. Sellers have been burned by tools with write access — a misconfigured repricer that tanked margins, an inventory sync that oversold a SKU. A status layer that cannot mutate anything is safe to leave running on a shared screen. That is a design principle worth demanding from your other tools.
One file, one HTML page, MIT-licensed. The maker notes it is MIT-licensed with themes as single files, and there are open “good first issue” tickets for new themes (hospital, newsroom, kitchen, classroom). For operators with a developer on staff, that is a weekend project to build a “warehouse” theme that maps to your actual fulfillment flow.
The Telegram control plane is the underrated part
Buried in the comments is the feature most sellers will actually use daily. The maker added Telegram commands: /mute WebDev silences an agent, /unmute restores it, /muted lists muted agents, /questions pings only when an agent needs you, /errors pings only on failed runs, and /all is both — the default. This is a notification-routing layer, and it is the part that maps directly onto how cross-border operators already work. If you have ever had a Shopify Flow automation spam your phone at 3 a.m. because a webhook fired, you understand why per-agent muting matters.
The maker also confirmed the mobile view shows “the same office data, in a layout that fits the screen,” with people who need you and errors staying visible and their cards sorted first. Tapping an agent opens a sheet with task, last steps, and on Paperclip the question plus its options. That is a genuinely useful pattern for sellers who are not at a desk — trade show, factory visit, 3PL walkthrough.
Where my judgment says it falls short
First, the aesthetic is a bet. Pixel-art offices with seven themes (pixel, holding HQ, plaza, warehouse, factory, space, military) will delight some operators and read as a toy to others. A category manager at a mid-market brand will not put a pixel office on the team’s main monitor. The maker leans into this — “Adorable :)” from Ryan Hoover is the top comment — but adorable is not the same as enterprise-ready.
Second, the integration surface is thin. Claude Code hooks are installed via npx @caglarutkuguler/cubicle install-hooks, which writes to ~/.claude/settings.json after backing it up. That is clean for Claude Code. Codex, Gemini CLI, and Paperclip are covered, but the source does not enumerate hooks for other runtimes. If you run a custom agent built on OpenAI’s API or Anthropic’s API directly, you are writing the JSON feed yourself.
Third, there is no persistence story described. “Replay” records a day and plays it back in minutes, or exports a 30-second video — useful for demos and postmortems. But there is no mention of long-term retention, search across historical runs, or alerting thresholds. For a seller trying to prove to a brand partner that a listing update ran correctly three weeks ago, that matters.
The real competitive risk
The incumbents are not standing still. Langfuse is open-source and could ship an ambient view. The CLI vendors themselves — Anthropic, OpenAI, Google — have every incentive to build first-party fleet dashboards. Cubicle’s moat is taste and speed, not technology. The MIT license and open theme tickets are a smart community play, but they also mean anyone can fork the office metaphor into their own product.
What I’d watch / test next
This week, if you run even two agents, do three things. First, run the live demo at caglarutkuguler.github.io/cubicle for ten minutes and ask yourself whether an ambient view would have caught a missed approval in the last month. Second, if you use Claude Code, run the install-hooks command on a non-production machine and verify the ~/.claude/settings.json backup before you trust it. Third, wire up the Telegram bot and set /errors as your default — failed runs are the alerts that cost sellers money, and question pings can wait.
Watch for two signals over the next quarter: whether the maker ships a rules screen (he calls it “the natural next step” — hide idle agents, show only errors), and whether a hosted, multi-machine aggregation layer appears. If both land, Cubicle stops being a cute side project and starts being infrastructure. If neither lands, it stays a delightful toy for solo operators — which, to be fair, is most of the cross-border seller market.
