The Real Bottleneck in Agentic E-Commerce Work Isn’t the Model — It’s Knowing When to Step In
Cross-border operators are quietly becoming agent herders. A single afternoon might involve a Claude Code session refactoring a feed-mapping script, a Codex CLI run generating localized listing copy, and a long-running scrape of competitor pricing on Temu and SHEIN — all while you’re on a supplier call or triaging an Amazon Seller Central case. The failure mode isn’t that the AI gets the task wrong. It’s that it stalls silently on a permission prompt or a clarifying question, and you don’t notice for twenty minutes. Evlat, a new macOS menu-bar utility from maker Omer Kala, is a small but telling answer to that exact problem — and worth dissecting because it points at where the whole agentic tooling stack is heading.
What Evlat actually is, stripped of the launch-day gloss
Evlat is a native Swift desktop app that lives as a thin strip on the edge of your Mac screen, rendering one ring per running agent session. When a session needs you — a permission request or a question — the ring turns amber. Hovering reveals every session plus your Claude and Codex usage limits. One click jumps to the right tab in iTerm, Warp, Claude Desktop, Metalterm, or Bateri. There’s also an evlat watch command that puts any long-running shell command — builds, tests, training runs — on the bar, and remote servers are supported over SSH. The app asks for no macOS permissions and keeps everything local. It’s free, and the code is on GitHub. “Evlat” is Turkish for “child,” which is either charming or ominous depending on how many sessions you’re running.
That’s the product. Now the interesting part is the architecture, because it tells you how fragile this whole category is.
How it differs from the incumbents you’re probably already using
If you’ve spent any time in the agentic tooling space, you’ve seen the usage-limit dashboards. Tools that track how much of your Claude or Codex quota you’ve burned are a dime a dozen, and several of them are genuinely good. Evlat’s maker explicitly frames the gap: “The tools I found were mostly about usage limits, not about what each session was actually doing.” That’s the wedge. Usage dashboards answer “how much have I spent?” Evlat answers “which of my six parallel runs is blocked on me right now?” For a cross-border operator juggling five workstreams at once, the second question is the one that costs money.
The comparison set matters. If you’re running agents inside a full IDE like Cursor or a terminal multiplexer like tmux, you already have some ambient awareness — a pane flashes, a tab title changes. Evlat’s bet is that the awareness layer should be OS-level and agent-agnostic, sitting above iTerm, Warp, Claude Desktop, and whatever terminal you happen to be in, rather than baked into any one of them. That’s a defensible thesis. It’s also the thesis that every menu-bar utility in history has tried and mostly failed to monetize.
Why Amazon sellers should care more than Shopify ones
Here’s where I’ll make a judgment call that the launch page doesn’t. The operators who need Evlat most aren’t the Shopify DTC crowd running a single Klaviyo-triggered flow. They’re the Amazon FBA brand owners and marketplace account managers running batch operations: bulk listing updates across hundreds of ASINs, scheduled repricing jobs, review-monitoring scrapes, ad-bid adjustments pushed through the Amazon Ads API. Those are exactly the long-running, permission-gated, easy-to-forget tasks that Evlat’s evlat watch command was built for. A Shopify merchant’s agentic workload tends to be shallower and more interactive. An Amazon operator’s workload looks like a queue of background jobs that occasionally need a human to approve a destructive action. That’s the amber-ring use case.
Where the math breaks
Two things in the launch thread should give any operator pause, and both are architectural.
First, Codex desktop sessions aren’t tracked. A commenter flagged it, and the maker confirmed: Evlat tracks Codex through its official hooks, and the Codex desktop app currently doesn’t fire hooks at all. The same ~/.codex/hooks.json fires every event from the Codex CLI but nothing from the app — a known upstream regression on OpenAI’s side. So if your workflow runs through the Codex desktop app rather than the CLI, Evlat is blind to it. That’s not Evlat’s fault, but it’s your problem.
Second, and more fundamentally, the detection mechanism is hook-dependent. The maker explained in the thread that Evlat uses hooks — Claude Code’s and Codex’s official extension point — with a tiny curl command POSTing each event to a loopback-only local server. “Waiting” comes straight from those events: a permission request or a question means the agent is blocked on you. Session files are only a supplement for session name and liveness, treated as estimates. The maker’s argument is that the core signal rides on the documented hook API, so if the session file format changes, “a detail goes quiet, but the core signal still rides.” That’s a reasonable position, but it means Evlat’s reliability is a function of how stable the hook APIs of two rapidly-moving vendors stay. If Anthropic or OpenAI changes their hook schema in a way that doesn’t break the CLI but does change event semantics, Evlat goes quiet without going down. You’d never know until you missed a prompt.
The SSH feature is the sleeper
The bit I didn’t expect, and the one most relevant to cross-border operators with infrastructure in multiple regions, is the SSH support. Remote sessions use the same hooks, and Evlat opens an ssh -R reverse tunnel so the server’s hooks report back to your Mac, each machine with its own key. If you’re running scrapers on a VPS in Singapore, a fulfillment-automation agent in Frankfurt, and a repricing job on a US box, that’s a single pane of glass for all three — without exposing anything to the public internet, since the tunnel is reverse and the local server is loopback-only. That’s genuinely useful for anyone whose “agent stack” is actually a distributed mess of cloud boxes. It’s also the feature most likely to be under-documented and under-supported six months from now, because reverse-tunnel setups are exactly where small open-source projects accumulate bug reports they can’t triage.
What cross-border sellers can borrow from this
Even if you never install Evlat, three patterns are worth stealing for your own operations:
- Treat “blocked on human” as a first-class state. Most agent orchestration treats success and failure as the two terminal states. Evlat’s whole design says the third state — waiting for you — is the one that actually costs you money. If you’re building internal tooling around Claude Code or Codex, log that state explicitly.
- Push awareness to the OS layer, not the app layer. The reason Evlat works across iTerm, Warp, Claude Desktop, and five other terminals is that it doesn’t live inside any of them. If you’re stitching together your own dashboards, resist the urge to build them into whichever tool you happen to use most.
- Keep the local-first, no-permissions constraint. Evlat asks for no macOS permissions and keeps everything local. For operators handling supplier PII, marketplace credentials, or customer data, that’s not a nice-to-have — it’s the difference between a tool you can run on a work machine and one you can’t.
Where my judgment says it falls short
The product is free and the code is on GitHub, which is great for adoption and terrible for sustainability. Free menu-bar utilities with a single maintainer have a well-documented half-life. The maker is clearly responsive — answering questions about hook architecture, SSH tunneling, and the Codex desktop regression within hours — but responsiveness isn’t a business model. If you build Evlat into your daily workflow, you’re making a bet on one person’s continued attention.
There’s also a scope question. The launch page shows a public roadmap with shipped and open requests, including “Add support for AGY and antigravity,” “Show Antigravity / AGY usage alongside Claude and Codex,” “Add option to include Codex/ChatGPT sessions,” and “Possibly allow approval right from the tab.” That last one is the most interesting and the most dangerous: approving a destructive agent action from a menu-bar ring is convenient right up until you fat-finger it. I’d want to see a confirmation step, an audit log, and a clear distinction between “open the session” and “approve the action” before I’d trust that feature in a production repricing or listing-update workflow.
And the Codex desktop gap is a real limitation for a meaningful chunk of users. If your team standardized on the Codex desktop app because it’s friendlier for non-engineers — which is exactly the profile of a cross-border ops hire — Evlat doesn’t help you today, and won’t until OpenAI fixes hooks upstream.
What I’d watch / test next
This week, if you’re running any agentic workload, do three things. First, install Evlat on one machine and run it against your actual Claude Code and Codex CLI sessions for a full working day — not a demo, a real day with real prompts — and count how many times the amber ring saves you from a stalled session. That number is your ROI. Second, if you have remote boxes running scrapers or repricing jobs, test the SSH reverse-tunnel setup on one non-critical server before you roll it out anywhere that touches live listings. Third, and most importantly, watch the OpenAI Codex hooks issue — if it gets fixed, Evlat’s coverage jumps meaningfully, and if it doesn’t, you should assume desktop-app users are permanently out of scope. The broader signal here isn’t Evlat specifically. It’s that the agentic tooling layer for e-commerce operators is being built right now, mostly by solo makers, mostly for free, and mostly on top of hook APIs that two very large companies control and can change without notice. Build your workflows accordingly.
