Hosting for the way you actually code now

You don't write Kubernetes manifests anymore. You don't schedule deploy windows. You describe what you want, your assistant writes the code, and you watch it work. That loop is the whole job now, and most hosting platforms were built for a different one.

They expect a separate dashboard for the deploy. They expect handwritten YAML, or a GitHub Action you wired up once and never want to look at again. They want a region, a build command, a buildpack, a Procfile. None of that is the work. The work is describing what you want and letting your assistant build it. The deploy step should fit inside that, not interrupt it.

Percher is the host that fits inside the loop. The first time you use it, that's the part that stands out. You finish a feature inside Claude Code, say "ship it", and a few seconds later the live URL is in the chat. No tab switch. No second login. No YAML. The assistant calls percher_publish as a tool, the build streams in the same conversation, and if anything goes wrong it calls percher_doctor and reads back the fix.

A typical session

Here's what it looks like from inside the assistant.

`` you : build a tiny URL shortener with PocketBase claude: (writes server.ts, package.json, schema) you : ship it to percher as "snip" claude: (calls percher_publish) → Building image (12s) → Health check ok → Live at https://snip.percher.run you : add a click count on each URL claude: (edits server.ts, calls percher_publish again) → Build cache: layer hit (8s) → Live you : the count isn't incrementing claude: (calls percher_logs --tail 50, reads error) → Missing migration. Apply it? you : yes claude: (fixes migration, calls percher_publish) → Live. Try it now. ``

The point isn't the speed, though it is fast. The point is that nothing in that loop required leaving the chat. The build log, the failure, the fix, and the redeploy all happen in the same place the code lives.

That last exchange is the one that matters. Your assistant will ship broken code sometimes. It guesses a package version wrong, forgets a migration, misses an env var. On a dashboard-first host that's where you become a courier: open a tab, find the deploy, copy the log, paste it back into the chat, wait. With Percher the assistant calls percher_doctor. Doctor reads the deploy logs, classifies the failure (build error, health check timeout, missing env var, OOM, crash loop), and returns a structured recovery.nextAction the assistant can act on directly: set the key, fix the line, retry, or roll back. The crash-to-fix loop closes inside the assistant.

What you get without configuring anything

A few things come standard, because a vibe coder shouldn't have to ask for them.

• Auto-detect. bunx percher init inspects your project and writes a working percher.toml. Next.js, SvelteKit, Astro, Vite, FastAPI, Django, Express, Bun: 29 frameworks detected out of the box. You don't pick a buildpack.
• PocketBase batteries. Any app can opt into a managed PocketBase instance: SQLite, authentication, and S3-compatible file storage in one container, reachable via POCKETBASE_URL. Set [data] mode = "pocketbase" in your toml and it comes up alongside the app. No separate database to provision, no auth provider to wire up.
• Outbound HTTPS proxy. Your app reaches any external API (OpenAI, Stripe, Resend, Polar, the Web Push services) by reading HTTPS_PROXY, which Bun and modern Node honor automatically. No host allowlist, no certificate dance.
• Predictable cost. €0, €3, €12, or €29 a month, flat. No per-GB bandwidth invoice on the weekend your post hit Hacker News.
• Rollback by tool call. "Roll back the last deploy" is percher_rollback with the previous deploy id. Instant, no manual steps.

None of this is Claude-only. Percher ships an MCP server that registers identically with Claude Code, Cursor, Windsurf, Continue.dev, or anything else implementing the spec. Run bunx percher mcp to print the config for your assistant. And if you haven't set up MCP at all, bunx percher publish works straight from the terminal. The CLI is the foundation; MCP is the polish that stops the context-switching. See the quick start for the two-minute version.

Where Percher fits, and where it doesn't

Percher is built for solo builders and personal apps. One region, in Germany. Flat per-account pricing. The opt-in PocketBase sidecar. That makes it a good fit for the things a vibe coder thinks of in the shower and ships by the afternoon: internal tools, side projects, weekend hacks, single-purpose utilities, the app you build for yourself, your family, a club, or a smaller business.

It isn't the right fit for a monetised SaaS or an app for paying customers. Percher today has no uptime SLA, no formal compliance certification like SOC 2 or HIPAA, and one region. That's a deliberate scope decision for the audience above, not a roadmap gap I'm hiding. A small operational model is the feature: one app, one container, one deploy, one health check, one rollback chain, something your assistant can hold in its head. But if you're shipping a multi-region SaaS, an app under a customer contract, or anything where downtime has financial consequences, a higher-tier provider is the safer call.

There's one place that scope shows up early. If you're coming from Vercel, edge functions need to be converted back to Node, since Percher's compute is regional containers, not edge. Image optimization works for Next.js apps unchanged. There's no direct equivalent to Vercel KV, Postgres, or Blob: use the PocketBase sidecar for the SQLite-plus-auth-plus-storage shape, or point DATABASE_URL at any external Postgres you already run. The comparison guides are explicit about where each competitor wins, including the ones where Percher loses.

That honesty is the point. The pitch isn't that Percher does everything. It's that for the apps you actually build through an assistant, the deploy stops being a separate step you leave the conversation for. You describe the thing, it ships, and when it breaks the fix happens in the same window. Free plan, no credit card. Then bunx percher mcp and you're shipping from the prompt.