Changelog · 2026
What shipped, why it shipped,
who benefits.
Mirox is built in public. Every entry below traces to a working part of the system you can read in the dashboard, the trace export, or the safety log. Reverse-chronological. No marketing dates.
- v0.9.1transparency
Founding-cohort Live is now Day 15 — fresh ASIN or your own campaigns, your call.
- What
- The 60-day founding-operator program is Shadow for two weeks, then Live from Day 15 onward. Scope sits in the dashboard — the seller decides what activates and when. Every live bid runs under the kill switch, Shield circuit breakers, Sentinel inventory throttle, and a written Day-21 incident protocol. Mirox is free for the 60-day program. Public Shadow Mode policy is unchanged — free, forever, opening to the public after the cohort wraps in Q3 2026.
- Why
- The earlier program left live bidding optional, which left half the codebase a paying customer touches — SP-API write path, kill switch under real pressure, Shield, post-bid telemetry, billing — un-exercised. We need real bids in production before public open. We briefly drafted a contract-enumerated scope (one marketplace, three campaigns, a daily cap) and dropped it the same day. The product already gives the seller per-campaign control. Scope belongs in the dashboard. The contract asks them to use Live in some form by Day 15. Nothing more.
- Who
- EU sellers spending €5,000+/month on Amazon ads who would rather help calibrate a transparent tool than wait for the polished version. Public visitors are unaffected — this is a private cohort policy.
- v0.9.0brand
One name. One product. One invoice line.
- What
- Sub-brands retired. Mirox is now the company and the product. Shadow Mode and Live Mode keep their names because they describe what the seller does, not what we sell. Master line locked: "The AI that shows its work."
- Why
- Sellers were asked to remember three names to evaluate one tool. That tax goes to zero. Sales calls get shorter, support tickets get clearer, and SEO concentrates on one canonical name from day one.
- Who
- Anyone evaluating us — fewer names to memorise. Cohort members — one application form, one workspace, one invoice line.
- v0.8.4feature
The inventory watch is now visible — bids step down before stockout, not after.
- What
- The system reads days-of-cover from your inventory feed and softens bids on ASINs about to run thin. When the feed goes quiet for more than an hour, bids cap at half — never blind. When inventory data drifts out of sync, every bid takes an automatic 50% haircut until it lines up again.
- Why
- Most PPC tools spend hardest on the ASIN that is about to stock out. The bill keeps running; the orders cannot. Sellers told us the average stockout costs them €18,000 in lost revenue and recovery work — so we stopped the wrong bid before it gets placed.
- Who
- Multi-marketplace sellers running thin cover. Anyone who has watched a Sunday-evening alert burn budget on an ASIN that already went OOS Friday.
- v0.8.0transparency
Every bid now exports as a row you can hand to a CFO.
- What
- One-click CSV from the dashboard. Each row carries the bid, the agent that proposed it, the predicted conversion rate, the relevance score, the alternatives ruled out, and the safety checks that were applied. Open it in Excel; defend it in the meeting.
- Why
- Sellers told us the same thing three times: "I cannot present a number I cannot explain." Most PPC tools answer "why did this bid change?" with a shrug. We answer with the full reason next to the number.
- Who
- Heads of marketing reporting to a CFO. Agencies producing client reports. Anyone tired of explaining "the algorithm decided" with no algorithm in sight.
- v0.7.5brand
Brand foundation locked — wordmark, palette, voice.
- What
- Inter 700 wordmark, tight tracking. Signal Blue #1D4ED8, Midnight #030712, Cloud #F3F4F6, Indigo #6366F1. JetBrains Mono for data and timestamps. Master line: "The AI that shows its work." Voice rules written down — no "revolutionary", no "leverage", no theatre.
- Why
- Sellers comparing six tools at once need to recognise us across the website, the cold email, the conference badge and the dashboard without re-translating. Locking the system once means we never have to second-guess a surface.
- Who
- Anyone who sees us — sellers, agencies, investors, programme committees.
- v0.7.0platform
Public pricing live — three tiers, no demo wall.
- What
- Three tiers in Stripe: Starter €149/mo, Growth €399/mo, Scale €899/mo. Plan limits enforced at the application layer. 14-day full trial with one rule — no live bids until the trust period is complete. End-of-trial transitions are automatic.
- Why
- Most PPC tools hide pricing behind a demo, then quote a percentage of ad spend. We do neither. The trial stays shadow-only because Live Mode needs a 30-day trust curve we will not pretend to short-circuit.
- Who
- Self-serve sellers comparing tools without booking a sales call. Founding-cohort members keep a bespoke €199/mo rate locked for 24 months.
- v0.6.4feature
Inventory-aware recommendations — bid less on what is about to run out, more on what is restocking.
- What
- Inbound-shipment-aware bid recommendations now surface alongside outbound bids on the dashboard. Recommendation only — auto-action stays a future tier. Each recommendation shows the days-of-cover number that shaped it, so the seller can sanity-check the math before clicking through.
- Why
- A PPC tool that cannot see your inventory bids hardest on the ASIN about to stock out. The mechanism had been running quietly in the background for weeks; what changed was making it visible to the seller, with the reason next to the number.
- Who
- Multi-marketplace sellers. Agencies running clients with thin cover. Anyone whose Sundays revolve around an inventory spreadsheet.
- v0.6.0feature
Better keyword matches in every language — relevance up 19%.
- What
- The model now reads product copy and search queries as different kinds of text — because they are. Similarity scores rose from 0.67 to 0.80 on production traffic. The cache rolled to a fresh version so old embeddings drop out cleanly.
- Why
- Most tools flatten the difference between how a seller writes a listing and how a buyer searches for it. Treating the two symmetrically is fast and wrong; treating them asymmetrically is the answer the literature predicts. The production data agreed.
- Who
- Multilingual EU sellers (the gain is largest where the source-language gap is sharpest). Anyone harvesting keywords at scale across DE/FR/IT/ES/UK.
- v0.5.0feature
Live Mode opens — for accounts that earned it in Shadow first.
- What
- Live Mode unlocks for accounts after 30+ days of clean Shadow Mode. All seven safety layers active. Kill switch is global and per-marketplace. Stale-data bids capped at half. Shadow Mode keeps running alongside Live for as long as the seller wants — both screens, side by side.
- Why
- Trust gets earned in shadow before it gets granted in live. Most tools ask sellers to take the leap on faith; we ask them to watch for two weeks first. The 30-day floor came out of the cohort kickoff calls, not out of a marketing meeting.
- Who
- Founding-cohort members who completed Shadow review and chose to go live.
- v0.4.5platform
Tenant isolation locked down — your account is yours, with two locks on the door.
- What
- Every database query carries a tenant id, and the database itself refuses cross-tenant reads. Application code adds a second guard on top. 472+ integration tests run against the lock. Audit logs verify that no cross-tenant access has occurred.
- Why
- Isolation is not a feature, it is a precondition. Agencies will lean on it explicitly when they arrive. Sellers should never have to ask whether their numbers can leak into another seller's dashboard — the answer has to be no, and provably so.
- Who
- Every seller, every day. Agencies will rely on it explicitly the moment they arrive.
- v0.4.0safety
The inventory watch went live — running in the background, ahead of being made visible.
- What
- Days-of-cover read directly from the inventory feed. Bids on thin-cover ASINs softened automatically. Rationing transitions skip the 24-hour cooldown when stock drops fast. When inventory data drifts out of sync, every bid takes a 50% haircut until it lines up again.
- Why
- Inventory awareness was the largest single missing safety layer. Until 2026-03-06 we honestly could not check that row green. Now we can.
- Who
- Internal — went live in the backend before it surfaced in the dashboard. Made visible to sellers on 2026-04-30.
- v0.3.5safety
Sensible bids from hour one — no two-week ramp on every new ASIN.
- What
- Category-aware priors mean new ASINs and new keywords start with reasonable bids on day one — no 2-to-4-week data ramp. When the inventory feed disappears entirely, a safe-cap kicks in and prevents blind aggressive bidding while the feed reconnects.
- Why
- Most PPC tools need 2 to 4 weeks of data before their algorithms produce sensible bids. That is a tax sellers pay on every product launch and every new keyword. The warm-start carries hour one; the safe-cap covers the failure mode.
- Who
- Sellers launching new ASINs (especially in vertical categories where category data is tight). Sellers expanding into a new EU marketplace where keyword data does not transfer.
- v0.3.0safety
Seven safety layers between the AI and your ad budget.
- What
- Global and per-marketplace kill switches. ROAS circuit breaker trips at 0.50. Spend anomaly trips at 4× baseline (6× JP, 3× BR/TR). Amazon API errors above 10% trigger a 5-minute auto-pause. Stale-data bids capped at half. A 14-day cooldown after any crisis trip. Marketplace-aware state transitions throughout.
- Why
- Live Mode requires the AI not to act as a single point of failure. Seven layers do not eliminate failure modes — they stratify them, so the worst case is a paused campaign instead of a runaway bill.
- Who
- Anyone running Live Mode. The safety layers are the reason we offer Live Mode at all.
What ships next
Roadmap is set by the cohort, not by a steering committee.
Founding-cohort members have a direct line to the team and influence the next quarter's release order. Public roadmap goes live once the first three case studies do.
No card · 10 seats · 24 months locked at the founding rate