Every few weeks someone asks me whether they should move their SEO into an always-on agent. They have watched a video where a bot on a $5 server writes forty articles overnight, and they want to know if it is real.
Half of it is real. The half that is real is not the half in the video.
An always-on agent is one that keeps working after you close the laptop. Hermes, OpenClaw and Grok Bot are the three people actually use. A chat agent is one you sit inside, like ChatGPT, Claude Code or Cursor. Both can do SEO. Both are useless at it until you give them live search data. The difference is not capability. It is whether you are in the room.
That data layer happens to be the same for both, which takes some pressure off the decision. ContextBolt SEO is a single MCP URL that works in Hermes, Grok Bot and OpenClaw exactly as it works in ChatGPT, Claude Code or Cursor, so live keyword volume, difficulty, SERPs and AI-visibility checks follow you whichever way you go, and all of it lands on the same board. You are picking a working style here, not locking yourself into a toolchain. Which style you need is the real question, and almost nobody frames it that way. So here is the honest version, including the part where most people asking for an always-on agent actually want a reminder.
- Chat agents are for deciding. You are present, you can correct a wrong turn in one sentence, and the feedback loop is seconds.
- Always-on agents are for remembering. Scheduled checks, weekly reports, monitoring you keep meaning to do and never do.
- Start with chat. Add an always-on agent when the thing failing is your calendar, not your thinking.
- Both need the same data layer. One MCP URL gives either kind live volume, difficulty, SERPs, and Search Console.
- The data is Ahrefs-grade, decision-useful and directional. ContextBolt SEO is $35/month, with a 7-day free trial.
What actually separates the two
Not intelligence. You can run the same model in both.
An always-on agent is a long-lived process with a scheduler and a message channel. It sits on a VPS or a hosted cloud, wakes on a cron or a heartbeat, does work nobody asked for in that moment, and pings you on Telegram, Slack or Discord. Hermes Agent from Nous Research is the clearest example. OpenClaw is the same shape with a different heritage, and Grok Bot is the same shape distributed through Cursor.
A chat agent is a session. It exists because you opened it. It has your full attention, your repo, your half-formed idea, and it stops the moment you stop typing.
The practical gap is bandwidth. In chat you can say “no, not that one, the one with the lower difficulty” and be understood immediately. An always-on agent gets one shot at interpreting an instruction you wrote on Tuesday, then acts on it at 9am on Friday while you are asleep. Everything good and bad about always-on agents follows from that one fact.
The jobs each one is actually good at
Here is the split I would defend.
| Job | Chat agent | Always-on agent |
|---|---|---|
| Picking what to write next | Strong. You argue with it. | Weak. It picks alone. |
| Weekly Search Console check | Weak. You forget. | Strong. It does not. |
| Auditing one page properly | Strong. Iterative. | Weak. One pass, no follow-up. |
| Competitor movement over months | Weak. No memory of last month. | Strong. That is the whole point. |
| Writing the actual page | Fine, with you editing. | Dangerous unsupervised. |
Read down the two columns and the pattern is obvious. Chat wins anything that needs judgment in the moment. Always-on wins anything that needs to happen on a date whether or not you feel like it.
The reason people get this wrong is that the second column looks less impressive. “It checks Search Console every Monday” is a boring sentence next to “it publishes while you sleep.” The boring one is the one that compounds.
Most people asking for this want a reminder
This is the uncomfortable bit, and it is the most useful thing in the post.
When someone tells me they want an always-on SEO agent, I ask what it would do on day one. Roughly four times in five the answer is a monitoring job. Watch my rankings. Tell me if something drops. Check whether the competitor published. Keep an eye on Search Console.
None of that needs an agent that writes. It needs an agent that remembers, has read access to real data, and can send a message. That is a much smaller, much safer thing to build, and you can run it for the cost of a coffee.
The fifth person genuinely wants a publisher, and that is where it goes wrong. Search does not grade your work for weeks or months. An agent publishing unsupervised is not a fast content engine. It is a slow, expensive mistake with a long feedback delay, which is the worst possible shape of mistake to automate. The SEO automation loops that hold up are lookups, comparisons and drafts. The ones that ruin sites are unsupervised writes.
So if you are shopping for an always-on agent, write down the day-one job first. If it is monitoring, you need far less than you think. If it is publishing, you need a human on the button no matter how good the agent is.
Neither one can see search data
Both kinds of agent share the same blind spot, and it catches people who assume the expensive setup solves it.
Ask any of them what to write about and you will get a list. The words will look like keywords. None of them will have a monthly volume, a difficulty score, or the live top ten sitting next to them. The model is guessing from training data that is months stale and never contained search volumes in the first place.
Browsing does not fix it. An agent can open Google and read the page, but Google does not print difficulty scores, and it does not show you the 900 other terms you should have considered. Reading one results page is not keyword research.
The Model Context Protocol, or MCP, is the standard fix. Anthropic published it in late 2024 as one way for agents to call outside tools. An SEO MCP server is that protocol pointed at search data, so the agent can pull volume, difficulty, live SERPs, competitor keywords, backlinks and your own Search Console.
The useful part is that it is the same server either way. You do not buy a chat version and a cloud version. One URL works in Claude Code and in a headless agent on a VPS, because both speak the same protocol.
What changes in the setup
The data is identical. The plumbing is not, and one detail catches everyone.
A chat agent runs on your machine, so a local server started with npx works fine. An always-on agent usually does not run on your machine. It runs on a VPS or in someone else’s cloud, and a localhost address there points at that box, not yours. A local MCP server is invisible to it.
That is why cloud agents need a public HTTPS endpoint, and it is the single most common reason someone’s Grok Bot or Hermes setup silently has no tools. The address was right for the machine you typed it on and wrong for the machine that runs the agent.
The per-client detail differs. Hermes takes a url key in ~/.hermes/config.yaml. OpenClaw takes openclaw mcp add with a Streamable HTTP transport. Grok Bot has no settings form at all, so you tell the bot to add the server in chat. Cursor’s own docs cover marketplace plugins rather than custom MCP, which is why that path is easy to miss; Cursor staff confirmed it on the forum on August 30, 2026. Full setup for each is on SEO for agents, and the per-agent walkthroughs are SEO with Hermes, SEO with Grok Bot and SEO with OpenClaw.
What one good scheduled job looks like
If you only ever run one always-on job, run this one. It is the drop watch, and it is the job that justifies the whole setup for most small sites.
Every weekday morning, the agent pulls the last 7 days of Search Console against the 7 days before that. It compares average position per query. It flags anything that fell more than 10 places and still has enough impressions to matter. Then it messages you the list, or says nothing at all if the list is empty.
Copy this prompt
Every weekday at 9am, pull my Search Console data for the last 7 days
and the 7 days before that. Compare average position per query. List
any query that dropped 10 or more positions and had at least 20
impressions. If nothing qualifies, send nothing. Do not write anything.Three details in there are doing the work, and each one comes from getting it wrong first.
The impressions floor. Without it you get twenty queries a day that moved from position 84 to 96, which is noise dressed as a report. Twenty impressions is a reasonable starting line, and you will want to raise it as the site grows.
The silence rule. An agent that reports “nothing to flag today” every morning trains you to ignore it inside a fortnight. A channel that only speaks when something is wrong stays worth reading. This is the single most-skipped instruction and it is the one that decides whether the setup survives a month.
No writing. The job ends at the list. The temptation is to have it draft a fix, and the moment you do that you have a publisher again, with all the risk that carries and none of the supervision.
That job costs nothing in credits, because Search Console tools read your own data. It is the cheapest useful thing an always-on agent can do for you, and it is the one people skip on the way to building a content mill.
What it costs to run each
Chat is the cheap one, because you already pay for it. If you have ChatGPT, Claude or Cursor, all you are adding is a data source. Nothing new to host, nothing to keep alive.
Always-on adds three line items. A box to run it on, which is $5 to $10 a month for a small VPS. A model bill, which now runs continuously rather than only when you are typing. And your attention, in the form of a channel that pings you. That last one is a real cost. An agent that messages you six times a day gets muted within a week, and a muted agent is worse than no agent because you believe it is working.
The data layer is the same price either way. ContextBolt SEO starts with a 7-day free trial, then $35 a month for 1,000 credits, and one URL works in both. Research tools cost one credit, backlink tools cost three, and Search Console costs nothing. Whatever either agent looks up saves to the same SEO Board, so the scheduled job and the chat session read from one history.
Size the schedule to the budget. A daily keyword sweep across 50 seeds is 50 credits a day, which is a month’s plan gone in three weeks. This is the most common way an always-on setup quietly stops working.
How to choose, in four questions
Answer these honestly and the decision makes itself.
Is the failing thing your thinking or your calendar? Thinking is chat. Calendar is always-on.
Would you read the output every time? If yes, you want chat, because you are going to be there anyway. If you would only read it when something looks wrong, you want a scheduled job with a threshold.
Does the job need to happen on a date? Weekly reports, monthly competitor checks and drop monitoring all do. One-off research does not.
Are you tempted to let it publish? If yes, stop and put an approval gate in before you set anything running. Not after. The gate is the difference between a worker and a liability.
Most people land on chat first and add one scheduled job later. That is the correct order, and it is also the cheap order, because you find out what you actually want monitored by doing it manually a few times.
What to expect from the setup
An always-on agent will not make you consistent. It will make the checking consistent. You still have to act on the message, and a Telegram ping you swipe away is exactly as useful as the report you never ran.
It is also not a rank tracker. None of these agents will email you because position 9 became position 11 on the one keyword you care about. Search Console plus a weekday schedule gets you close, on your own queries only, and ContextBolt SEO’s daily Search Console snapshots give that job a history to compare against.
And the data is Ahrefs-grade, not Ahrefs. The estimates are decision-useful and directionally right, good enough to pick a target and walk away from a losing SERP. They are estimates, as every third-party index’s are, so pair them with Search Console when you need your own real numbers.
The version that works is unglamorous. One agent you sit in for the thinking. One scheduled job that watches the things you forget. Live data under both. And a human who still presses publish.
That is a worse video and a better quarter.
The data layer under both is one URL. ContextBolt SEO starts with a seven-day free trial, so the drop watch above can be running before you pay for anything.
