Zerops

🎉 BIG ANNOUNCEMENT: ZCP is live! A cloud dev platform where your coding agent works alongside you in the same workspace. Bring your own agent and subscription, run a full MCP-native dev loop, and keep production clean as you build. Details below.

@mustaphaDevFS Hey, there is a missing option: Zerops. Real Linux containers plus managed Postgres and Valkey on a private network, billed per minute of hardware.

Where do you deploy in 2026? A) Vercel B) Railway C) Fly .io D) Cloudflare Pages

@eiranoirx I have a new solution for you, and it's Zerops. Real Linux containers, managed Postgres on a private network, hardware pricing per minute.

which hosting u actually stuck with long term → Vercel → Railway → Render → AWS → Heroku → DigitalOcean → Home server (mad respect) what didnt disappoint u

@LazyCode3 Well, Zerops skips the free-tier funnel entirely: credits to start, then you pay for the cores and RAM you use by the minute. No pricing cliff to plan an escape from.

Free tiers are not free. They are customer acquisition funnels with an expiry date. Supabase pauses your project after inactivity. PlanetScale killed their free tier overnight. Heroku removed free hosting entirely. Always have an exit plan.

@Akintola_steve Worth a look before you settle: Zerops. Real Linux containers with managed Postgres on a private network, billed per minute of hardware rather than per service. Same ease, more control when the project grows.

The moment you migrate away from render to railway you’ll have peace of mind. Or are there am better deployment platforms you’ve come across that one can play around with ?

@coltcoanmsft @doncampbell @OmarShahine Hosting it as a container is the right instinct! In case you're interested, on Zerops that MCP server runs as a real Linux container on a private network, so your agents reach it without it being exposed to anyone else.

You can make content creation as easy as you’d want. For my site, I had it make an admin portal that uses a Zero Trust access policy in Cloudflare for authentication. It has a rich post creation suite. The other thing I did, was have it create a MCP server that I host as a container app so I can have my agents create posts on my behalf, very easily. It’s a new world!

Done! Site is now alive on Cloudflare. Thanks Fable! "You're welcome, Omar! Fun project — 30 years of blog history safely redirected, a fresh landing page, and Squarespace on the clock. Just remember to pull the plug on that subscription before January 2027 (I'll remember too). 🎉"

@Tushkiz The microservices parallel holds, and it extends to the network layer. On Zerops the MCP servers share a project private network, so auth and reach are bounded by topology even before a gateway, not hard-coded per integration.

Running MCP without a gateway is the same problem we solved a long time ago for microservices. In dev, direct connections are fine. In prod, every capability you skip the gateway on becomes a decision you make once per server, not once per system. Without a gateway, every connection point means: - Auth is each server's problem - Rate limiting is per-server, or absent - Audit logs are scattered or missing - Policy enforcement is hard-coded per integration - Observability you build from scratch, N times With 2 MCP servers, this is manageable. With 8, it is a maintenance problem. With 20, it is an incident waiting to happen. The gateway is not just a security add-on. It gives the system one place to enforce shared rules.

@Plaksha_ Credentials in env plus a reachable MCP is the dangerous combination. If you're interested, on Zerops the server runs on a private network reaching only what you wired in, so even a compromised one has no route to the systems that credential unlocks. zerops.io/zcp

When an agent signs a payment transaction, it uses a credential it holds. If that credential lives in an env file and most do,then a compromised dependency, a prompt injection attack, or a misconfigured MCP server can now initiate financial transactions on your behalf.

x402 just made AI agent security 10x more urgent. Here's why I think most teams building on it aren't thinking about this. 🧵 @x402

@JDSalbego Hey! The hidden connections are the scary ones, because you can't audit what you don't know you have. Network scope makes them harmless: on Zerops an MCP server only reaches its project's private network, so a forgotten install still can't touch anything outside it.

The average AI agent workspace has more MCP connections than the developer realizes. Most builders install 3-5 intentionally. The actual count is usually higher. Where the extras come from: 🔵 Bundled dependencies: you installed one MCP server, it pulled in two more as dependencies 🔵 IDE defaults: your editor ships with built-in MCP connections you never configured 🔵 Framework scaffolding: your agent framework pre-configures servers at setup time 🔵 Forgotten installs: servers you added three months ago for one task and never removed Each one is a connection to an external service with access to your resources. Each one is an attack surface. 🔵 82% of MCP implementations are vulnerable to path traversal 🔵 41% have zero authentication 🔵 67 CVEs filed in 2026 alone Run the audit. List every server. Remove what you don't need. The connection you forgot about is the one an attacker will use.

@KnowTechGlobal If anyone is interested, on Zerops the deploy path is one yaml in your repo and the containers are standard Linux, so nothing about your workflow is trapped behind a proprietary model you can't leave.

On Nov 28, 2022, Heroku's free dyno stopped being the default parking spot for small apps. The choice looked simple: pay, move, archive, or vanish. But the real loss was not $5 or $7. It was a whole deploy path hiding behind one command.

@Bhawarthaa You forgot Zerops! Real Linux containers with managed Postgres on a private network, billed per minute of hardware rather than per instance

Which cloud provider do you deploy to most? • AWS • Google Cloud • Azure • Vercel / Railway / Render Curious which most developers are deploying with.

@debba_92 @__spekulator__ @Sn0wbrave Just wanted to mention that on Zerops each project is its own private network, so client A's server has no route to client B's database at all - maybe someone will find it useful!

@__spekulator__ @Sn0wbrave yep . MCP server needs per-project schema isolation or the agent will happily hallucinate table layouts from client A onto client B's DB.

🚀 Working on something cool? Drop your GitHub project in the comments. Include: • Repository link • A short description • What problem it solves Let’s build a thread full of amazing open-source projects. You might discover your next favorite tool.

@pauliusztin_ Hey Paul, in case you're interested, on Zerops it's a container with a stable internal hostname on the private network, so swapping Claude Code for Codex changes nothing underneath.

Your AI assistant shouldn't care which harness you're using. Claude Code today. Codex tomorrow. OpenCode next month. The only thing that should move is one configuration. Your memory, tools, and skills should come with you. This is why I like to think of the MCP server as the portability layer for AI assistants.

@barbinbrad @zachglabman @SutterlinDan @c0nst @will_alverson If we may also add something: On Zerops the backend and its MCP server share a private network, so agents reach it directly while it stays off the public internet.

what's up zach! 🫡 couple scattered thoughts: - from the beginning of time (the 70s), manufacturers have hoped software would be a silver bullet, and it never has been - i think excel/whiteboards is a good solution for smaller shops - sendcutsend may eat the parts market with software while everyone is waiting for AI to help them - SAP has taken the approach of locking down APIs, and keeping the agentic experiences INSIDE of SAP - our thought, currently, is the exact opposite. we have no "AI" inside carbon. instead we expose our entire backend as MCP server, so claude can use it to do stuff. the cost of this is visibility into what people are doing, but it's a better experience imo. - we want people to use LLMs to modify carbon itself or build on top of it. - speculative, but imo: this is the future of in-product agents (everyone get's their own openclaw agent essentially that writes code, makes integrations, and does stuff proactively): x.com/_overment/stat…

@SutterlinDan many schools of thought cc @c0nst @will_alverson @barbinbrad

@waqasdaha Solid setup for a painful migration.... On Zerops that MCP server runs beside Vertica on a private network, so Claude reaches both raw and transformed data without either being publicly exposed.

True, I’m currently deep in one of the most complex pipeline migrations from Airflow DAGs to real-time @ApacheFlink + Java. After many painful verification iterations on upper envs,I built an MCP server that lets Claude pull both raw data and the transformed data from Vertica.1/n

@velddev Nice self-healing loop, Mike. When you want it running beyond a slow PC, Zerops gives the daemon and its MCP server a real container on a private network, so Codex reaches them without a public endpoint.

My pc was really slow today, so i ended up making an metrics aggregator daemon with an mcp server so codex can get notified and fix it

@NOWNodes Exactly, and once it's a tool call rather than context, where the server runs starts to matter. On Zerops it sits on the same private network as the data it exposes, so the call never leaves the project

Why an MCP server instead of pasting docs into your prompt? Docs in context go stale, eat tokens every turn, and the model still has to find the right method in a wall of text. MCP makes it a tool call. The assistant asks for the one thing it needs, when it needs it.

The NOWNodes MCP Server is live. Instead of digging through API docs, сonnect #Cursor, #Claude, VS Code, or any compatible AI coding assistant to our #MCPServer. Generate production-ready API requests based on the latest NOWNodes documentation, saving up to 30% of tokens compared to standard AI workflows. $BTC, $ETH, $BSC, $TON and and 120+ networks of what you actually use. No extra cost, MCP included in your plan. Here's how it works:

@evan_tarv Hey Ecan, ten minutes to a public URL is great until the public URL is the problem... In case you're interested, on Zerops the same stateless server deploys just as fast but onto a private network, so your agents reach it without it being reachable by everyone.

4/ Part 1: Scaffold the server. A TypeScript MCP server with a stateless /mcp endpoint, wired to the SDK. Claude Code writes it from one prompt. Push it, then deploy to Railway for a public URL with zero DevOps overhead. Done in ~10 minutes.

1/ A customer just used your app without ever opening it. They asked their agent instead. 🧵

@sundi133 Hey Jyotirmoy, pinning and diffing definitions catches the change. Containment catches the intent. On Zerops a server acting on a poisoned description still only reaches its project network, so the injection has nowhere to send anything.

Reminder: an MCP tool's description is part of the prompt. A compromised server can hide instructions in the tool's name, description, or parameter docs. Your model reads all of that before it ever decides to call the tool. Pin tool definitions. Diff them on every update.

@cdatasoftware Cross-server exfiltration is the one containment answers directly. On Zerops each MCP server sits on a private network reaching only what you wired in, so a coerced client has no route to the next server's data. Thanks for this thread.

These aren't theoretical. The advisory documents real incidents: tool parameter injection via unsanitized inputs, cross-server exfiltration through coerced MCP clients, GitHub MCP granting blanket repo access, output poisoning across chained agents.

MCP has no required authentication. That's a protocol-level design choice. Authentication is left to implementors. Many skip it entirely. The NSA's May 2026 advisory documents what happens next. 👇

@4A4556494C Hey, I think we mentioned it before, but on Zerops the server's reach is the project's private network, so even inheriting the agent's full privilege set it can only touch what that network reaches

1/ MCP servers are being actively scanned in the wild. I've covered the recon angle. What I haven't said clearly enough is that port exposure is the least interesting part of this story. The real issue is transitive privilege. An MCP server doesn't have its own permissions. It inherits the full privilege set of the LLM agent it serves — every API key, every file system path, every tool integration the agent can invoke. Compromise the MCP endpoint and you don't get a dumb service. You get a proxy with the agent's entire capability surface.