How persistent memory works, what you can automate, what an agent costs to run, and how the tools and hosts compare. Written by the team that builds Hivra.
Cursor renamed Background Agents to Cloud Agents, and people keep asking the same two things. What does it cost? And does it keep going when I shut my laptop? We went through Cursor's own docs. Here's what they actually say.
Read the articleClosing the laptop on a coding agent: what happens to it, which fixes work, and how to leave one running safely.
Closing the lid stops a long Claude Code run, and some tasks in the Claude app too. Below: what locking, sleep and a closed lid each do to Claude, then five fixes that keep Claude Code going while you're away, from a one-line command to a computer of its own.
Remote Control lets you steer a Claude Code session from your phone or a browser. Claude keeps running on your machine, and code execution and file access stay there. When that machine sleeps, loses its network or has Claude Code closed, people report "Can’t reach your computer" or "Claude Code on the computer running this session is offline".
OpenAI announced reusable Codex cloud environments at DevDay on 29 September 2026 and is rolling them out gradually, so for some jobs you don't need a computer of your own any more. The table below puts it next to the Codex CLI on a computer that stays on, your laptop and a rented VPS.
Plenty of Claude Code vs Codex comparisons test them as assistants you sit beside in a terminal. More and more people run them another way: on a server, around the clock, with nobody watching. Judge them like that and the picks change.
Most Cursor vs Claude Code posts crown a winner on vibes. We're not doing that. The useful question is simpler. Do you want to sit and watch the agent edit, or do you need it to keep working after you leave? We went through Cursor's and Anthropic's docs. No speed race, and no "smarter model" verdict, because we haven't timed either one on your code.
Claude Code, Codex, OpenClaw, Hermes: every terminal AI agent dies when you close the terminal and stalls when the laptop sleeps. The fixes run from a 90 second tmux setup up to systemd, and past those there's an always-on computer at $9.99/month.
Running an agent 24/7 comes down to four decisions. Where the process lives, what restarts it when it crashes, what it remembers, and how it reaches you at 3am.
Anthropic's official Telegram channel plugin lets Claude Code receive messages from Telegram. You pair your own account, and the session keeps running where it already runs, on your computer or server. If that machine sleeps or Claude Code exits, the bot stops receiving work.
Leaving an AI agent running while you sleep is safe for most workloads, as long as four things are under your control first: permissions, spend, isolation and recovery. Each one has a mitigation that works.
One flag turns off Claude Code's permission prompts, and Anthropic says to use it only inside a container or VM. We went through Anthropic's and OpenAI's docs and checked Hivra's defaults against our own code. Short version: a few checks survive the flag. A separate computer keeps a mistake off your laptop. Anything you sign in to on it is still fair game.
Codex stores each chat on disk, so you can leave and come back. We tested every command below on Codex CLI 0.159.0. Claude Code's resume flags are further down, since that's a different tool.
Your own server, a VPS or a managed computer. How to pick between them, what each needs, and what an agent costs to run.
Search for AI agent hosting and you'll find enterprise cloud docs and thin listicles. Neither tells you where a personal or CLI agent should live so it keeps working after you close your laptop. There are four options, with prices for each.
In cash, self-hosting an AI agent and paying for managed hosting sit a few dollars a month apart. Hours are where they split: setup, updates, security, and who gets paged when the agent dies at 3am.
Most AI agents are API orchestrators that idle on 1 vCPU and 2GB of RAM. Sizing the server is the easy bit. Keeping a machine that holds your API keys and logins safe on the public internet is the hard part.
Hermes Agent runs well on cheap cloud servers, and at €5.49/month you can afford one. Pick a tier that covers your usage, from a provider the community trusts. Go too small and things break.
AI sounds like GPU territory, so people ask. For an agent that calls a model through Anthropic, OpenAI, Google, OpenRouter or another cloud API, though, the provider runs that model on its own hardware. Your machine just runs the agent, its tools and any browser or code processes it starts.
An agent that runs all the time costs more than the price on the plan. Here's each piece with a real figure: the server, the tokens, your time, and the bits no pricing table mentions.
Claude Code is included in Pro ($20 a month), Max 5x ($100) and Max 20x ($200), prices as of September 2026. Anthropic gives Max's extra usage only as a multiple of Pro's per five-hour session. It doesn't currently publish message or token counts. So you pick by where Pro stops you. The rule for that is below, along with the break-even against API billing, what to do at a limit, and when Anthropic's cloud is enough.
OpenAI folds Codex into ChatGPT plans, and the plan cards keep moving. So we went back through the pricing page, the Pro tiers help article and the speed docs and put it all in one table. One rule never moves: a computer of your own does not raise your Codex allowance.
Cursor renamed Background Agents to Cloud Agents, and people keep asking the same two things. What does it cost? And does it keep going when I shut my laptop? We went through Cursor's own docs. Here's what they actually say.
Ask what an agent costs per month and you usually get a list of tips. You wanted a bill. Here's one, on Anthropic and OpenAI prices as of October 2026, plus the few cuts that actually move the number.
BYO means bring your own: you supply the AI provider credentials your agent uses. Requests go straight from the agent to the model provider, and the provider sends you the bill. Hivra isn't in that loop.
What Hermes Agent is, how to run it yourself, and how to use its memory, skills, schedules and chat apps.
Nous Research released Hermes Agent, an open-source autonomous AI agent, in early 2026. A chatbot answers when you ask and then stops. Hermes keeps going between your messages.
Hermes Agent runs on a VPS, keeps its memory and talks over Telegram while you sleep. Getting there from a blank Ubuntu server still takes a few hours when things go right. Docker shows up in two different ways in Nous's docs, and mixing them up is how people lose a weekend. Hivra is one managed path. It is not Nous, and it does not run Hermes on My cloud or My server.
If you know Linux, self-hosting Hermes Agent isn't hard. The trouble starts later. Browser automation, memory volumes and updates are where most setups bite back, and for some people it just isn't worth it.
Hermes Agent runs 24/7 on a server, so it keeps working after you close the chat. Before you set one up, you'll want to know what you'd use it for. These are the jobs people in the community give it, with enough detail to tell whether any of them fit you.
Skills are Hermes Agent's version of plugins: reusable capabilities the agent can call on demand. Unlike other mainstream agent frameworks, Hermes writes skills for itself. When it solves a problem it'll face again, it packages the solution into a skill and improves it over later uses.
Most AI tools forget everything the moment you close the window. Hermes holds memory in four overlapping places, each with its own job. What sticks across sessions depends on which layer caught it. Some things still fall through.
A persistent agent doesn't wait for you to ask. Hermes fires tasks at whatever interval you define: every morning, every hour, once a week. You don't need to know cron, though the syntax works if you do.
Hermes Agent receives messages and replies on 15 platforms: Telegram, Discord, Slack, WhatsApp, Signal, SMS, Email, Home Assistant, Mattermost, Matrix, DingTalk, Feishu/Lark, WeCom, Open WebUI, and Webhooks. They all handle security the same way. Without an allowlist, anyone who finds your bot can message it, so set one before you expose a gateway publicly.
Telegram is one of the easiest ways to talk to a Hermes agent. You message the bot from your phone like you'd message anyone, and the agent replies. Setup is a short list of steps.
People often ask: 'Why would I use Hermes Agent when I already pay for ChatGPT?' Both are AI, and both write and code. They split on the kind of work each one can do.
How to run OpenClaw and Agent Zero, what to do when one breaks, and how they compare with Hermes.
OpenClaw is one of the most-used open-source AI agent frameworks, with a large community and plenty of integrations. Getting it running on a VPS takes 30-90 minutes if you follow these steps. Keeping it stable, updated and secure is the harder part.
OpenClaw ships new releases several times a month, and some of them break things. Agent stopped responding, lost its permissions, or won't start after an update? Usually it's a changed config schema, and that's the first fix below.
Agent Zero and OpenClaw are two of the most popular open-source agents you can self-host. They're built for different jobs. They ask different things of a server and cost different amounts to host, and either one needs protecting once it's exposed to the internet.
Hermes Agent and OpenClaw both let you run an autonomous AI agent on your own server. Both appeared in 2025-2026 and grew fast. They're built differently, though, and which one suits you depends on what you want the agent to do.
Start here if agents are new to you. What an agent is, how it remembers, what it can automate, and where several agents fit.
People call everything an 'AI agent', from a chatbot with a search button to a system running around the clock on company servers. The technical meaning is actually pretty clear. Pin it down and the rest follows: how agents work, where they earn their keep, and where the hype runs ahead of them.
Give an AI agent a task and it starts a loop. It keeps going until the work is done. Here's what happens at each step, from the first model call to the finished result.
Plenty of marketing copy uses 'chatbot' and 'agent' as if they meant the same thing, and they don't. The difference decides which one to reach for.
Start a new chat and the model has forgotten the last one. Persistent memory is what carries over. It matters for any task that spans more than one session.
A 200K token context window sounds huge. Then you run an agent every day for a few weeks and find out it isn't storage at all. Memory has to live somewhere else, and a handful of systems now do that job.
Multi-agent systems went from research demos to real deployments in 2025. Gartner expects 40% of enterprise AI deployments to use them by 2028. So which framework fits which job, and what does it cost you?
Specific examples are the best way to tell whether a persistent AI agent is useful for your work. Here are seven tasks that run well on a schedule. Each one comes with the setup it needs.
Browser work is the most useful thing an agent can do, and the thing people most often set up wrong. The four main tools differ a lot on success rate and cost per task. They also fail in very different ways.
