Meltbox GTM · launch kit

The Show HN launch kit.

A Show HN is the one $0 event that behaves like a paid traffic spike for people who run coding agents. It is also a one-shot consumable. This kit exists so the shot is taken with real data — not survivorship stories — and only once the traffic has somewhere good to land.

The honest numbers

Mined from the true Show HN population over the last 3 years via the HN Algolia API (empty-query population pull, no relevance bias). N = 79,893 Show HN; agent-tooling comparable cohort N = 9,483. Re-mine before any launch date — node scripts/shn-miner.mjs.

CohortNMedianP75P90P95<10 pts100+ pts
All Show HN79,89324124788%3%
Developer tooling14,93023103190%2%
Agent tooling (our set)9,4832492390%2%
Read this plainly: the median Show HN gets 2 points and goes nowhere. 88% never clear 10 points. The front page is a lottery. And the agent-tooling cohort is tougher than average (P95 of 23 vs 47) — the category is saturated, so standing out is harder here, not easier. A "good" launch in this category is the ~150–310 band, not thousands.

The real comparables (live points)

Comparable launchPointsCommentsDate
Omnara — Run Claude Code from anywhere category ceiling3101682025-08
Vibe Kanban — manage your AI coding agents1951322025-07
OneCLI — Vault for AI Agents161522026-03
Agent Vault — credential proxy for agents156552026-04
Boxes.dev — ditch localhost; run CC in the cloud105792026-06
Pulse — approve tool calls from your phone cautionary39132026-06

The cautionary one matters most. Pulse — "Dashboard for Claude Code, approve tool calls from your phone" — got 39 points, and its top comments were "Claude Code already does this natively." Lesson baked into the titles below: do not lead on phone-approval, and lead on what no comparable does — the agent building you a decision, not a log dump. Separately, "Continue? Y/N — a 60-second game about AI agent permission fatigue" scored 386 points: permission fatigue is a proven HN nerve, and it's exactly Meltbox's wedge.

The headline decision: the window

Not the content — the timing. This is decided in the brief; here's the reasoning.

Recommended: fire after the demo + AEO are live on a real domain recommended

A Show HN is one-shot. Today the traffic would land on either an app login wall (meltbox.ai root is the app, not a landing page) or a noindex preview. Even a median Show HN with a real landing page beats a front-page hit that bounces. Two cheap prerequisites unlock the shot: (1) the docs robots.txt AI-crawler block is lifted (the AEO fix), and (2) the hello-demo page is live on a real URL. Both are other GTM items already in flight.

also on the table  this week (fire now, accept the login-wall bounce) · after realtime ships (wait for a bigger story, risk the momentum cooling). The brief lets Will pick.

5 title variants — each carries its data

Winner shapes from the mined top set: Name – concrete what-it-does (em-dash) dominates; concrete numbers over-index; first-person "I built/used…" recurs; 0% of top titles use marketing adjectives. All five obey that.

1  “Meltbox — A decision inbox for the coding agents you already run” recommended

Data: em-dash winner shape (Omnara/Semble/Mastra). "Decision inbox" is concrete and names the one mechanic no comparable has; "you already run" targets the exact ICP. Sidesteps the phone-approval trap that capped Pulse at 39.

risk "inbox" could read as email-adjacent — the first comment must clarify fast.

2  “Meltbox — A control plane for the coding agents you already run”

Data: the product's own line; precise ICP targeting. Clean, no adjectives.

risk "control plane" scans slightly enterprise/abstract next to the concrete winners.

3  “Meltbox — Stop babysitting agent approvals; get pinged only when it's irreversible”

Data: rides the permission-fatigue nerve that scored 386 points, and echoes a real operator on the Pulse thread: "phone only pings when something's irreversible… per-call approval you stop reading by day two." Positions against Pulse/native remote-control instead of competing with it.

risk "babysitting" is mildly loaded; two clauses.

4  “Meltbox — I was drowning in agent output across terminals, so I built a review inbox”

Data: first-person origin framing recurs among winners ("I built a sub-500ms voice agent", "I put an AI agent on a $7 VPS"). Humanizes; leads with the shared pain.

risk longest of the five; some HN readers want the tool named first.

5  “Meltbox — Let agents use your API keys without ever seeing them”

Data: the secret-proxy framing pulled 156–161 points for Agent Vault and OneCLI — proven live demand for exactly this sentence. Concrete and specific.

risk foregrounds one pillar and hides the brief story; invites the "prompt-injected agent still exfiltrates through the proxy" objection (answered in the crib sheet).

Post body & founder first comment

Drafts. The post body is the submission text; the first comment goes up within two minutes of posting. Both honest, concrete, ending in a real question — HN rewards a genuine ask.

POST BODY

Meltbox is a control plane for the coding agents you already run.

The bottleneck stopped being "can the agent do it" and became "the agent
produces decisions faster than I can review them, scattered across terminal
tabs, localhost pages, and Slack DMs to myself." So the agent hands me a
*brief* instead: a small clickable HTML page with the actual choice laid out.
I triage them in an inbox with keyboard shortcuts (j/k to move, e to archive),
and my pick + note flow straight back to the agent as its next step. When
something actually needs me, my phone gets a push — Mac and iPhone, even with
the app closed.

It also proxies your API keys: an agent gets a scoped key that resolves to the
real token server-side, so it can *use* a connector without ever *seeing* the
token. And it gives agents one authed place to build and update internal
dashboards against shared data.

Onboarding: sign up (email, Google, or anonymous guest -> workspace instantly),
then hand a one-time invite link to Claude Code / Codex / any LLM. It redeems a
scoped, revocable token and says hello with a brief.

It's built by its own briefs — the PWA push, the MCP hub, keyboard triage, and
the ChatGPT app were all shipped by an agent proposing a decision brief and me
picking. Build log (8 real rounds): [build-log link]

Free plan, works today at meltbox.ai. The thing I most want feedback on: is a
clickable *page* the right unit for an agent to hand you a decision — or do you
just want a better notification with a link back to the terminal?
FOUNDER FIRST COMMENT (post within 2 min)

Founder here. Origin: I run a lot of agents, and "three tabs of Claude Code, two
localhost pages, and a Slack DM to myself" became the actual job. I built the
thing I wanted — agents come to me with a decision I can act on in two
keystrokes, and my phone pings only when it's my turn.

The part I'm least sure about and most want to argue about: I made the atomic
unit a *brief* — a functional HTML page the agent builds — instead of a
notification + a link back to a terminal. The bet is "here are 3 options, pick
one, leave a note" beats "come look at my logs." The whole product is built this
way: [build-log link] is 8 real rounds (PWA push, MCP hub, ChatGPT app,
realtime) where an agent proposed a brief and I clicked.

Stack: single Cloudflare Worker, D1, Durable Objects for realtime presence,
Web Push (VAPID) for the phone. Secrets are a proxy, so the token never enters
the agent's context.

Happy to answer anything — especially "why not Slack / ntfy / the built-in CC
remote control." Short answers below; poke holes.

Objection crib sheet — real quotes, real answers

Every objection below is a verbatim comment from a comparable launch thread (Pulse, Omnara, Vibe Kanban, Agent Vault, OneCLI, ntfy). Answer in your own words on the day — never paste.

“Claude Code already does this natively.”

“Claude Code's built in remote control feature already lets you respond to permission prompts from the Claude app on your phone.”— m12k, cebert, felooboolooomba on the Pulse thread (this was the #1 objection)
Answer: Right, and if you just want to approve a bash command from your phone, use the native feature. Meltbox isn't per-call approval — a real operator on that same thread said it best: "per-call approval you stop reading by day two." It's a decision surface: the agent hands you a structured choice with options and context when it hits a fork, it works across agents (Codex, Cursor, cron), not just Claude Code, and the same inbox holds your keys and dashboards.

“No moat — I'll roll my own in an afternoon.”

“why should a hacker submit to using tools behind a SaaS offering… when one can just roll their own tools? I don't think there is any sort of moat here.”— mccoyb on Omnara
“it's just… a pub/sub server… using HTTP? … something most devs can hack in an hour or two.”— brap on ntfy
Answer: You can absolutely roll your own — ntfy + a Slack webhook + a .env + a kanban board gets you 70% there (there's a comparison page showing that exact stack working). What you don't get for free: the decision loop where your click becomes the agent's next input, one key vault every agent shares, and closed-app phone push, in one authed place, maintained. Free plan — you're not paying to try it. Honest: if you enjoy maintaining the glue, the glue works.

“You've reinvented SSH + tmux.”

“So you've reinvented SSH+screen except slower and much less flexible?”— globular-toast on Omnara
Answer: For remote access, ssh+tmux wins and stays local. Meltbox is the opposite: it exists so you don't attach to the session. The agent comes to you with a decision; you're not driving a terminal from your phone.

“Where do my keys and data live? Does it see my code?”

“I like that vibetunnel keeps all of my data local. Does this have remote access to my codebase and session?”— rgbrgb on Omnara
Answer: Meltbox never needs your codebase — the agent runs on your machine/VM; Meltbox holds the briefs, the app data you put there, and proxied secrets. On secrets: the proxy resolves a scoped key to the real token server-side, so the token never enters the agent's context or logs.

“The proxy doesn't actually stop exfiltration.”

“a prompt-injected agent can still exfiltrate data it reads through the proxy. The trust boundary shifts, it doesn't disappear.”— c5huracan on OneCLI; echoed by paxys, hanyiwang on Agent Vault
Answer: Correct, and we don't claim otherwise. The proxy closes key leakage — the token never lands in agent-readable memory, logs, or a screenshot — and lets you scope + revoke per agent in-band. It does not stop a compromised agent from using the access it was granted; nothing at that layer can. That's a blast-radius problem, which is why access is per-agent and revocable. We state the boundary plainly rather than overselling.

“Why these permissions? Are you harvesting my data?”

“The permissions this asks for feel kinda insane… Why does a kanban board need to see the code or my deploy keys?” — and analytics defaulting on, harvesting email + GitHub usernames.— doritosfan84 & gpm on Vibe Kanban
Answer: Meltbox holds only what you hand it — briefs, app data, connectors you add. Analytics is first-party and product-event only, with no default email/username harvesting; we shipped it that way on purpose (build-log round 8) after watching launches get roasted for the opposite.

“What happens if you shut down?”

“If you don't trust me or your messages are sensitive, run your own server. This is the way.”— krono on ntfy (the answer HN respects)
Answer: Your agents keep running — they run on your machine, not ours; Meltbox is the control surface, not the runtime. Briefs and app data export out. Killing Meltbox doesn't kill your agents — that's the honest mitigation, not a longevity pinky-promise.

Launch-day run-of-show (Pacific time)

The single hard requirement: Will is present the entire day to reply to comments. A Show HN with an absent founder dies. And never ask for upvotes anywhere (Slack, X, DMs) — HN detects vote rings and will bury or ban the post.
  1. Re-mine + freeze. Run shn-miner.mjs for fresh numbers, lock the title, confirm the gate is met (robots lifted, demo page on a real domain, both fast).
  2. Prep, don't prime. First comment + crib sheet in open tabs. No coordinated upvote outreach — none.
  3. Post. Submit the Show HN with the frozen title + short body (morning PT catches the US workday; Show HN doesn't need midnight). Post the founder first comment within 2 minutes.
  4. Reply to everything, fast. This is the whole game. Answer objections from the crib sheet in your own words. Watch the demo page's first-party analytics (won't be ad-blocked).
  5. Stay in it. Lean into the "why not X" debates honestly — HN rewards a founder who argues in good faith. Do not repost or edit the title.
  6. Momentum check. Climbing → stay glued. Flat (the likely case — median is 2 points) → that's fine and expected; keep answering, don't delete.
  7. Second channel + learn. Run the Product Hunt listing (different audience) if the window fits. Thank real feedback. Log recurring objections into the crib sheet for the next round.

Product Hunt listing (secondary channel)

Launch Tue–Thu, 12:01am PT. Assets: logo 240×240 PNG, 3–5 gallery images, demo GIF. Tagline <60 chars, no jargon, no adjectives.

Tagline options

“Your coding agents hand you decisions, not log dumps” 52 chars
“A decision inbox for the coding agents you run” 46 chars

DESCRIPTION (≈220 words)

Meltbox is a control plane for the coding agents you already run — Claude Code,
Codex, Cursor, cron jobs, anything. Instead of scattering their output across
terminal tabs, localhost pages, and Slack DMs, your agents hand you *briefs*:
small clickable pages that lay out a decision — "here are 3 options, pick one,
leave a note." You triage them in a keyboard-fast inbox, and your choice flows
straight back to the agent as its next step. When something actually needs you,
your phone pings — Mac and iPhone, even with the app closed.

Three more pieces come in the same authed home: a Secret Proxy so agents can
*use* your API keys without ever *seeing* the tokens (scoped and revocable per
agent); self-building Apps so agents can spin up and update internal dashboards
against shared data; and a remote MCP server that works in ChatGPT.

Onboarding takes a minute: sign up (email, Google, or anonymous guest), then
hand a one-time invite link to your agent. It redeems a scoped token and says
hello with a brief. Free plan.

Meltbox is built by its own briefs — its PWA push, MCP hub, and ChatGPT app all
shipped by an agent proposing a decision and a human clicking approve.

Gallery shot list

1 · Brief Inbox with the keyboard-triage overlay (j/k/e)  ·  2 · a real brief — 3 options + per-item note box  ·  3 · iPhone lock screen: "New brief in Meltbox"  ·  4 · terminal redeeming an invite → "says hello with a brief"  ·  5 · Secret Proxy diagram (agent → scoped key → real token, server-side)  ·  6 · the build-log: "shipped through its own briefs"

Social copy decks (drafts — nothing posts)

X / THREAD (5 posts)

1/ I run a lot of coding agents. The bottleneck isn't the agents anymore — it's
that they produce decisions faster than I can review them, scattered across
terminal tabs, localhost pages, and Slack DMs to myself. So I built Meltbox.

2/ The idea: your agent hands you a *brief* — a small clickable page with the
actual choice — instead of "come look at my logs." You triage an inbox with j/k
keys. Your pick flows back to the agent as its next step.

3/ When something actually needs you, your phone pings — Mac + iPhone, even with
the app closed. Not per-call approval (you stop reading those by day two). Only
the decisions that are yours.

4/ Keys: agents get a scoped proxy key that resolves to the real token
server-side. The agent can *use* a connector without ever *seeing* the token.
Revoke per agent.

5/ It's built by its own briefs — PWA push, the MCP hub, the ChatGPT app all
shipped by an agent proposing a decision and me clicking. Free plan, works
today: meltbox.ai
LINKEDIN (single post)

The hard part of running coding agents in 2026 isn't getting them to do the
work — it's that they produce decisions faster than one human can review, and
the output scatters across terminals, localhost pages, and DMs.

Meltbox is the control plane for that. Your agents hand you decisions as small
clickable pages — pick an option, leave a note, and it flows back to the agent
as its next step. Your phone pings only when a decision is actually yours. Keys
are proxied so agents can use connectors without ever seeing the tokens.

It's built by its own briefs. Free plan at meltbox.ai — feedback welcome.

The proof asset

The first comment's credibility rests on this: 8 real shipped rounds, each proposed as a decision brief and chosen by a human.

Open the build log →