The team
Which agents you see depends on what your project is. A business that never approaches people, such as a law firm or a team inside a big company, gets no Outreach or Creator agent, and its Reddit agent only listens.
The month’s plan decides how much of this work your business gets. For the Reddit, SEO, Trend, Listening, Directories, Applications and Creator agents, it picks whether each one runs, how often (never more often than the table says) and how many tokens it may use until your tokens renew, with one line on why. An agent your business has no use for stays off, and its card says Off this month with the plan’s reason. One that is on is funded for its whole month at the rhythm the plan picked: a quarter of your account’s tokens for the month is kept for the agents of all your running projects, and each gets what a month of its runs costs, or more where the plan wrote more. When they ask for more than that quarter holds, it is shared out fairly, and an agent whose runs cost more than its share can carry runs less often and says so on its card, such as “Runs about every 3 days this month, not daily”. On Solo, each agent gets its share of the plan’s own tokens, divided among your running projects. The Reddit, Trend and Listening agents share one line of the plan, each with its own part of its tokens. An agent’s month is spread over its days: one that has used today’s part says Resting on its card, and one that has used its whole month says Waiting, each with the day it runs again. Being ahead of the month’s pace never stops an agent. You see and change those choices on Plan, under Agents this month. An agent whose work nobody uses stops and tells you: UpSell stops looking for threads, MarKet stops writing trend ideas, and MarList and MarPly prepare no new form or application, until you or your own AI use one. See When something is blocked.
Marsell can also ask one of these agents for one extra run before its turn, when the next days depend on fresher findings, such as the day before a launch. Its daily pass or the chat asks, with a reason you read in the activity feed: “Marsell ran Discovery early:” and why. It only happens while that agent is on and is neither resting nor waiting, the run spends from its tokens for the month, and an agent never has two early runs at once. An early run does the agent’s usual work: discovery reads the fresh conversations even inside a day of its last read; AI answer checks run a full pass without waiting the week; listening reads Reddit, the trends and the web for pages that talk about you; creator search looks again, once you’ve started it; press and directories look for directories and “best of” lists; site checks look at your site, your products and prices, and your live pages again; and video readings read each video with a new week of watch time. When there’s nothing new to do, the same line in the activity feed says so, and why, such as “Nobody posted anything new in your communities on Reddit since the last read.”
Each card is titled with the agent’s own name, with what it is (such as “Reddit agent”) under it. It shows what the agent does, the platforms it works on, how often it runs, what it has done lately, and when it last ran. Press a card, or See what it did, to go where its work is. When something it made is ready for you to see, the card shows Review with how many, and opens them. The Creator agent waits until you start it: its card says “Not started yet” and Start it opens Market’s Creators tab. The Directories agent’s card lists the places to list you, best fit first, with Copy for your agent (see Get listed in directories). MarList and MarPly are twins: their two cards share one frame, side by side, wherever the team is shown.
When an agent still needs something from you, such as a postal address for the foot of its emails, a booking link or the name it signs with, its card says so in one quiet line: “Needs 2 things from you”, or “Could use 1 more thing” when it can work without them. Press it to open what that agent needs from you, such as What WarSell needs from you: first what the agent can’t work without, then what would help, then what Marsell already has, each with where it came from. Each one saves on its own, right there; what Marsell has can be changed, and taken off where that makes sense. An account to connect, a password or anything that spends is never typed here: the row says where in Marsell you do it, with a button that takes you there. Marsell’s chat and your own AI read and fill the same list; the details only MarPly reads, such as the name you apply under, show their value only to people who can approve applications, in Marsell. A list, such as subreddits, searches or topics, holds up to 20 and is saved whole. A card with nothing missing shows no line. What each agent may ask for:
- UpSell, MarKet and MarBell: subreddits you want read, topics your buyers care about, and other names to search for (a nickname, a misspelling, a product line).
- MarCo: questions your buyers ask an AI assistant, your site for his pages, and Google Search Console.
- MarList: your website, your privacy policy and help pages, who makes it and where it’s based (as your listings show them), your logo and pictures for listings, the email people write to, your phone number and your social accounts.
- MarPly: who applies, the name you apply under, when it started, the country, money raised, who backs it, its stage, the team, the tools it uses, what you’d spend money you win on, and pictures for applications. These stay private.
- MarStar: what the creators post about, the fewest and most followers, where the creators are and where they post.
- WarSell: who signs the emails, your postal address, your booking link, your yes to finding and emailing people, and the mailboxes to send from.
- Ben: where customers leave you a review, your shop or checkout, how long an order takes to arrive, and your review request email.
- MarMail: a way for people to join your list, an email address to send from, people on your list, a welcome email, and your postal address.
Set how an agent works
An agent whose work you can steer has one switch: On its own, Show me each or Paused. Agents start On its own: on a project that goes ahead on its own, their work goes within its limits from the start; otherwise each kind of work they do learns from your first few answers, then runs by itself within its limits. Either way it’s logged and undoable. Show me each keeps every piece coming to you, and anything it had already lined up to go comes back to you as a card. Paused stops it, and nothing it made goes out: anything already waiting to go is held, listed first under What I do next on Home as Held, until you switch the agent back, and you can still undo it meanwhile. It costs nothing while paused. What you write, send or approve yourself is never held by an agent’s switch: it goes at once (see What you press yourself goes at once). Pausing the Outreach agent here also pauses its campaign at the mailboxes, so its follow-ups wait too, and switching it back starts the campaign again; pressing Start on Sell’s Outreach tab switches it back here as well. Pausing the Creator agent here turns it off under Agents this month on Plan too, and turning it off there pauses it here. Press a setting to save it for that agent. The Reddit agent has all three. The Creator and Outreach agents can only be paused or running, so Show me each is greyed out on their cards (“Not for this agent”). The other agents have no switch yet, so their cards show none. How the switch works explains it. Money is different: every spend still comes to you as its own card, and so do offers to creators. See What needs you.The Reddit agent
The Reddit agent reads the subreddits that fit you, reads each thread worth answering in full, and hands you the posts from the last three days where you can genuinely help, with a reply written when you ask for one. It never answers threads about health, legal trouble or grief, and never sends DMs. It reads each subreddit’s own rules first: where they don’t let a business take part, it hands you no thread there. A reply only tells what your business has really done, from your facts, your own posts and videos, or your notes. A reply that names you says that you run the business, in its own words and the thread’s language, or in the exact words you gave Marsell for it. Before it writes, it reads what the poster has said to you before, on any channel, and it never gives away in public what they told you in private. Its card, titled “UpSell, your Reddit agent”, shows how it’s doing (Healthy, Needs attention, Blocked, Starting or Paused), its checks, its last run (what it read and found) and when it runs next. It needs at least three checked subreddits, and Marsell checks more each week. How it works explains it in plain words, and after a run What it did says what the last one did.- Without Reddit connected (“Public tier · it finds threads, you answer them”), it reads every hour while the month’s plan runs listening daily or more often, and you post the replies yourself.
- With Reddit connected (“Reddit connected · posts the replies it has learned”), it can read as often as every 15 minutes. The first replies you ask for in each subreddit come to you; once it has learned them, it posts a reply you asked for that doesn’t name you on its own, with a day to undo. Set it to Show me each and every reply comes to you, the ones already waiting their day included. Pause it and a reply already waiting is held until you switch it back. A reply in someone else’s thread is Marsell approaching them, so on a project that goes ahead on its own it doesn’t learn: each one comes to you, unless replies there already ran on their own before, or you set them to.
The accounts they work through
The accounts they work through, “What each one can do today, and what it can’t yet”, lists each kind of account, marked Connected, Not connected or Needs setup, with What it can do to read it in full:- Social accounts post to your own accounts, learning from your first few posts on each.
- Reddit finds threads and writes replies in your voice; connected, it posts the replies it has learned from you.
- LinkedIn posts to your LinkedIn.
1
Connect Reddit
Under The accounts they work through, press Connect next to Reddit, then Continue on Reddit.
2
Allow access on Reddit
Approve Marsell on Reddit’s own page.
3
Check it took
Back in Agents, press Check accounts. Reddit should read Connected.
Your own AI
At the bottom of Agents, Your own AI lets Claude, ChatGPT, Claude Code, Cursor, OpenCode, Codex or any agent you use work on the project with Marsell, through Marsell’s MCP server. It sees what you see, and answers what waits for you only when you ask it to. Work you’d otherwise do by hand can go to it in one paste. On What needs you, a stack of directory listings, replies to post in communities, messages to send or replies to paste on social has Copy for your agent: one brief with each one’s exact words, where it goes and the account it goes from. Your AI does each one from your own accounts and reports it back to Marsell with the link to what it posted.Connect Claude or ChatGPT
At the top, Connect Claude or ChatGPT needs no key. Press Copy next to Marsell’s address, then follow the three steps for your app, side by side on the card:1
Open the app's connectors
In Claude: Customize, then Connectors, then + and Add custom connector. In ChatGPT on the web: Plugins, then +, and add a custom server. Each has a link that opens the right page.
2
Paste the address
Name it Marsell and paste the address.
3
Sign in and press Allow
The app opens Marsell. Sign in if it asks, check what it may do (Can act is on; switch it off to let it only look), and press Allow.
Homework for your AI
Some work needs your own logins, files or inbox, where no connection of Marsell’s reaches. Homework for your AI is one list of all of it, whichever agent it came from: posts for accounts Marsell can’t post to (YouTube posts and polls among them), the Reddit threads UpSell found (each reply is written when your AI takes its thread, on your tokens), your directory listings as one batch, each application to a grant or program you said yes to (MarPly’s, from your own browser on the program’s site), the missions, real photos of a place from your own files, jobs Marsell’s plan writes (YouTube Studio work, replies to reviews on sites with no way in for Marsell, event listings, edits in your own site builder), and any card you hand over with Give it to my AI in What needs you. Your AI picks them up over Marsell’s connection, does them and reports back, with a link for each thing it did or why it left one out. A note your AI adds for you shows where that work’s result shows: on the job’s line, the reply’s line or the listings’ line. Each job says where it is done, and your AI does it only from there:- Your missions profile: a separate browser profile on your computer, signed in only to Reddit, X and LinkedIn, never to your email or bank. Everything that goes on Reddit, X or LinkedIn in your name goes from it, and nothing else is done from it. Your AI says it is in that profile each time it takes such work, and is refused otherwise.
- Your mailbox: an email you handed over goes from your own mailbox, as you, or through a mail connector of your AI’s that sends from it. A draft isn’t sent: an AI that can only make a draft leaves it there and gives the job back.
- Your files: real photos, and the end card your films end on, from your computer or a files connector of your AI’s.
- Your own accounts: your own browser, signed in to the site the job names, such as YouTube, a directory or Google Ads.
- Give my AI homework, at the top right of the card, turns the list on. Nothing goes on it until an AI that can act is connected: an app above, or a key with Can act on. While it’s on, Marsell’s plan can give your AI whole jobs, each one written out with every word it will post or send.
- Auto mode, in the panel at the foot of the card, lets your AI do Marsell’s planned homework without a card for you first. With it off, each planned job comes to you in What needs you as “Homework for your AI”, with the whole brief, and Give it to my AI or Skip. Only you can answer that card, in Marsell: your AI can’t say yes to its own homework.
- Beside auto mode, always shown, is a warning: use it only with a newer, careful model, such as Claude Opus. Marsell’s plan reads outside web pages, and an older or less careful model could be tricked into posting or sending something you never saw. You are responsible for what your AI does in your accounts.
- “Wakes on its own” with how often, then what it set, whether it wakes even while its app is closed or only while it’s open, and until when, if it stops by itself.
- “Its timer runs older instructions”, if your AI set it with an earlier prompt of Marsell’s. It still works. Bring its timer up to date, in the panel at the foot of the card, has a line to paste into your AI that puts the new prompt in that same timer, never a second one; if your AI can’t find or change it, it tells you and changes nothing.
- “Not on a timer”, with why, if you said no or its app can’t schedule. It checks when you ask it: say “check Marsell’s homework”.
- “Its timer stopped on” that day, if it did. The next time it reads its homework, it asks you about a new one.
- “Not on a timer yet”, if it hasn’t said. The next time it reads its homework, it asks you.
- “Last checked its homework” and when, or “Hasn’t checked its homework yet”.
Missions for your AI
Where you said yes to your AI sending on Reddit, X or LinkedIn (in setup, or in Settings, under Who Marsell talks to for you), Marsell hands it missions: single messages Marsell writes and checks, which your own AI sends word for word from your missions profile. Marsell sets the pace, keeps to each platform’s limits and tells you what went out. Missions go out from 9 in the morning to 9 at night in your saved time zone, never on Sundays, and at half the limits in a platform’s first week.
On LinkedIn, Marsell asks to connect with the people it keeps whose LinkedIn profile it has. When it knows something real about someone, such as what they sent your business or did there, the request carries a short note written for them, up to five notes in any 30 days (fewer in a platform’s first week or after a warning). Everyone else gets a request with no note. When someone accepts, Marsell writes them a first message and your AI sends it. If Marsell holds that message back, a card in What needs you says why, with Marsell’s draft under it, and the message you write there goes through your AI.
On Reddit, when your AI posted Marsell’s reply to someone who asked in public, Marsell may write that person one private message for your AI to send: the day after the reply went, or once they answer it, within a week of the reply. Never a second one, and never to someone who asked not to be contacted, who wrote to you by message since (Marsell answers them there), or who was approached in the last 30 days.
Missions for your AI, under Your own AI, shows each platform as a tile: whether it’s on (your answer to who Marsell talks to, with Change it), what went out today, what waits, and anything resting after a warning or a refusal, with Resume. The latest missions are rows under it; open one to see the message and where it went.
- Make a missions key, in the panel at the foot of the card, gives a key that can only take and report missions, never touch the rest of your business. Give it to an AI that works in a separate browser profile, signed in only to Reddit, X and LinkedIn, never to your email or bank. It shows once.
- Pause missions, at the top right of the card, stops them at once: your AI gets no missions until you press Resume missions, which needs a role that can approve.
- Missions go out through the day, so your AI should wake every hour. Wake it every hour, in the same panel, has a line to paste into it that sets that up, and says when your AI last asked for a mission.
- Your AI never rewords a message to get it through. A message it wouldn’t or couldn’t send comes to you as a card ready to copy, at most five a day, and two refusals of the same kind on one platform within a week stop that kind until you resume it.
- What people send back in the threads and conversations missions started, your AI passes on to Marsell from your missions profile before its next mission there; Marsell answers each as it does every reply, and the answer comes back as a mission. A pass with nothing in it loses nothing: the next one still looks back to the last reply it passed on, and a reply passed twice is kept once.
- A warning or a rate limit on a platform halves its limits for a week. If no AI of yours has asked for its missions while they wait (“Is its computer on?”), or a platform warns or restricts an account, Marsell tells you in the morning report, and right away on a linked Telegram or Slack. That message means nobody read at all; an AI that reads its list but doesn’t send gets the card above instead.
Apps and keys
Each app you connect shows under Connected apps and keys, with when it was connected, when it was last used and what it may do. Disconnect stops it on its next call.Connect with a key
For Claude Code, Codex or any other agent:1
Make a key
Name the key, for example “Claude on my laptop”. Switch Can act on to let it answer what you ask it to, make pictures and spend tokens, or leave it off to let it only read. Press Make a key.
2
Copy it
The key shows once: copy it, then press I’ve copied it.
3
Connect your agent
Under Connect with a key, pick Claude Code, Codex or Any MCP client and copy the setup it shows, with your new key already in it.
claude mcp add --transport http marsell with Marsell’s address, then type /mcp in Claude Code, pick marsell, choose Authenticate, sign in and press Allow. If it says marsell already exists, run claude mcp remove marsell first, then the line again. Cursor and OpenCode sign in with no key too: see Marsell MCP server.
Connected apps and keys lists the keys in use; Revoke turns one off. See Marsell MCP server.
From your own files
Under it, From your own files shows an example of your own AI, on your computer, setting up a project in Marsell from the files you already have: the brand guide, the logo and the fonts. Copy the prompt copies what to ask it. Your AI fills the setup; you check it in Marsell and confirm, and Marsell starts.What needs you
What Marsell brings you, and how it learns.
Marsell MCP server
Connect your own AI.