Melete

The Agent That Never Logs Off Is Coming for Your Back Office — Here’s What Cross-Border Sellers Should Actually Do About It

Every cross-border operator I know is running the same ragged playbook: a supplier thread in WeChat, a refund queue in Amazon Seller Central, a TikTok Shop comment pile, a Klaviyo flow that needs a rewrite, a 3PL integration that broke at 2 a.m. Shenzhen time. The bottleneck was never intelligence — it was continuity. You explain context to a chatbot, get a decent answer, close the tab, and the context dies. That’s why Melete, a new always-on agent from maker Yashas Gunderia, is worth more than a cursory look from anyone running multi-marketplace operations. It’s a bet that the agent should persist, own a browser, and survive your laptop closing — which is exactly the shape of work a seven-figure seller does every day.

What Melete Actually Solves — and Why It’s Not Just Another Chat Wrapper

The maker’s own framing is blunt and, frankly, correct: “My to-do list was full of things AI could do, and none of them got done. The best AI tools answered questions brilliantly, then left all the doing to us.” That’s the gap. ChatGPT, Claude, and Gemini are phenomenal at reasoning inside a session and useless at carrying a task across days, tools, and model switches. Melete’s pitch is an agent with its own computer, its own browser, connected to your email and calendar, running ²⁴⁄₇, picking work back up when a reply, a date, or an event arrives.

Three design choices matter for operators:

• Model-agnostic by default. You can run Claude, GPT, Gemini, or open models, switchable at any time. For sellers, that’s not ideology — it’s cost control. Route routine classification (review sentiment, tracking-number extraction, supplier email triage) to a cheap open model, and reserve frontier models for negotiation drafts or listing copy.
• Approval gates with receipts. The agent asks before anything that matters goes out, and leaves a receipt for everything it does. For anyone touching buyer messages, supplier POs, or ad spend, this is the difference between a useful tool and a liability.
• Inspectable memory. Melete shows you what it has learned about you and lets you correct it. That’s the feature most agents bury, and it’s the one that determines whether you trust the thing with a real workflow.

Open source, but read the fine print

The GitHub repo is live at ychampion/melete, and the maker clarified in the thread that the open-source release is the core infra — the whole stack — which you can run locally, while cloud is “the managed runtime for people who don’t want to self host.” A commenter pushed on exactly this ambiguity, asking whether the GitHub version is the full thing or a smaller slice of what the cloud will be. The honest answer: self-hosting gets you the engine, not the managed sandboxes, VM routing, and model orchestration that Melete Cloud will sell. For a seller with a technical co-founder or a dev on retainer, that’s fine. For a solo Amazon operator, it’s a hard pass until the hosted version ships.

How It Stacks Up Against What You’re Already Paying For

Let’s be specific, because “AI agent” is now a meaningless category.

Versus ChatGPT and Claude directly. You already use them. The failure mode isn’t quality, it’s amnesia. Every new thread is a cold start, and every model upgrade resets your accumulated context. Melete’s answer — when a better model ships, you switch, and your agent and its history come with you — is the single most operator-relevant line in the entire launch. If you’ve ever rebuilt a supplier-negotiation prompt from scratch because you migrated from GPT-4 to a newer model, you know the tax.

Versus Zapier and Make. These are deterministic automation. Great for “when order ships, send tracking email.” Useless for “chase this supplier about the QC photos, escalate if no reply in 48 hours, and draft a backup PO to the second factory.” Melete is aiming at the judgment layer, not the plumbing layer. In practice you’d run both.

Versus Helium 10, Jungle Scout, and the Amazon tooling stack. These are vertical tools with narrow surfaces — keyword research, rank tracking, PPC bid management. They don’t touch your inbox, your supplier WhatsApp, or your returns inbox. Melete is horizontal. The risk of horizontal is that it’s mediocre at everything; the upside is that it can stitch together the six tools you’re already paying for.

Why Amazon sellers should care more than Shopify ones

Shopify DTC operators live in a relatively clean stack — Shopify admin, Klaviyo, a helpdesk, maybe Recharge. The workflows are API-friendly and well-documented. Amazon sellers live in a hostile environment: Seller Central has no real API for half the things you need, suspension appeals are email-driven, supplier comms happen on WhatsApp and WeChat, and TikTok Shop and Temu add two more fragmented dashboards on top. That mess is precisely where a browser-owning agent earns its keep. An agent that can log into Seller Central as you, pull the stranded-inventory report, and draft the removal order is worth more than one that writes product descriptions.

Where the math breaks

Two cost lines kill naive agent deployments. First, token burn. A ²⁴⁄₇ agent that polls, reasons, and drafts is not a $20/month ChatGPT subscription — it’s a metered inference bill that scales with activity. Model-agnostic routing helps, but you need a hard budget cap per workflow or you’ll wake up to a four-figure API invoice. Second, error cost. An agent that misreads a supplier’s MOQ and auto-sends a PO is worse than no agent. Melete’s approval gate mitigates this, but only if you configure it conservatively for the first 30 days. I’d rather approve 200 actions in month one than clean up one bad one.

What Cross-Border Sellers Can Steal From This Launch

You don’t have to adopt Melete to learn from how it’s positioned. Three transferable ideas:

1. Memory is the moat, not intelligence. Every seller I know has institutional knowledge trapped in someone’s head — which factory actually delivers on time, which SKU has a 4% return rate for sizing, which ad angle worked in Germany but flopped in the UK. If you’re building any internal AI workflow, the priority is a persistent, editable knowledge store, not a smarter model. Melete’s rewindable memory roadmap is the right direction; you can replicate a crude version with a structured Notion or Airtable base your prompts always load.

2. Approval gates are a feature, not friction. The maker explicitly built asking before anything that matters goes out into the core loop. Copy that pattern into your own automations. Any workflow that touches buyer-facing communication, pricing, or supplier commitments should require a human click. Everything else can run free.

3. Model-agnostic beats model-loyal. The commenter Oleksii Sekundant nailed the appeal: “Being able to switch models and keep the project’s memory is the detail I like here. Re-explaining the same project every time gets old quickly.” If you’re locked into one vendor’s memory format, you’ve handed them pricing power over your own operational history.

The first jobs I’d hand an always-on agent

Ranked by ROI for a cross-border seller:

• Supplier follow-up chaser. Every PO, every QC request, every sample shipment — tracked, chased, escalated. This is pure calendar-and-email work that eats 5–10 hours a week.
• Marketplace health monitor. Poll for policy warnings, listing suppressions, and stranded inventory across Amazon, TikTok Shop, and eBay, then draft the fix and wait for approval.
• Review and returns triage. Classify incoming reviews and return reasons, flag product-quality signals, draft responses in the right locale.
• Ad-spend watchdog. Not bid management — anomaly detection. “Spend on this campaign doubled overnight, here’s the likely cause, approve the pause?”

None of these are exotic. All of them are currently done badly, late, or not at all by lean teams.

Where My Judgment Says It Falls Short

I’ll be direct, because the launch thread was polite and you don’t need polite.

The cloud isn’t live. Melete Cloud is open for waitlist signups, and B2B pilots are being taken at melete.si. That means today, the only thing you can actually install is the self-hosted core — which the maker admits is early, with bugs still being fixed. For an operator whose business runs on uptime, “early and buggy” is a disqualifier for anything customer-facing. Sandbox it on internal tasks only.

Security posture is the elephant. You are handing an agent your email, your calendar, and eventually — per the roadmap — your own laptop and browser as the agent’s home base, so it acts as you on the sites you’re already logged into. That is a catastrophic blast radius if anything goes wrong. The promised privacy layer that keeps account numbers and health details from ever reaching a cloud model is exactly the right instinct, but it’s roadmap, not shipping code. Until it exists, I would not connect a primary Amazon Seller Central account. Use a sub-user with scoped permissions, or a secondary email, and treat the agent like a new VA on probation.

The B2B pitch is thin. “Melete is an AI workspace for teams” is a tagline, not a product spec. Multi-seat permissioning, audit trails per user, and SSO are table stakes for any team deployment, and none are detailed in the launch. If you’re a 20-person agency running stores for clients, wait for the pilot terms before committing.

The competitive window is closing fast. The maker opens by referencing OpenAI’s launch of always-on agents with their own computer and browser a day earlier. That’s the real threat: platform-native agents from OpenAI, Anthropic, and Google will eventually do 80% of this with zero setup. Melete’s defensible ground is openness, model-agnosticism, and self-hosting — which appeals to a specific operator profile, not the mass market. That’s a viable business, but it’s a niche one, and you should price your dependency accordingly.

The self-host question nobody answered cleanly

The back-and-forth between Diae Louali and the maker is the most useful part of the thread. Louali asked whether the GitHub version is the full thing or a subset, and got a straight answer: the open-source is the core infra which you can run locally, while cloud adds managed routing, stateful persistent VMs, and secure sandboxes. Translation for sellers: if you self-host, you own the ops burden — model keys, VM hygiene, backups, and security patching. That’s a real cost, and it’s the reason most of you will end up on the waitlist rather than the repo.

What I’d Watch / Test Next

This week, before you commit a single workflow:

1. Clone the repo and run it air-gapped. Point it at a throwaway Gmail account and a Google Calendar with two fake events. Watch how it handles a multi-day task with a mid-task interruption. If it loses state, you’ve learned everything you need to know in an afternoon.
2. Join the cloud waitlist at waitlist.melete.si and, if you have a team use case, apply for a pilot at melete.si. Early access terms matter more than features right now.
3. Write down your top three recurring back-office tasks and estimate hours per week. If the total is under five, the ROI isn’t there yet — wait. If it’s fifteen-plus, start designing the approval gates now.
4. Audit your credentials. Before any agent touches your stack, create scoped sub-accounts for Amazon Seller Central, Shopify, and your email. Never hand over a root login.
5. Set a token budget. Decide your monthly inference ceiling before you turn anything on, and pick your default model accordingly.

The always-on agent is coming whether or not Melete wins. The operators who benefit will be the ones who treat it like a new hire — scoped access, tight approvals, measured ramp — rather than a magic button.