Connect Claude or ChatGPT
Marsell’s address ishttps://getmarsell.com/mcp. In the app, Copy puts it on your clipboard: it’s in Connect your AI at Start, and at the top of Your own AI on Agents.
- Claude
- ChatGPT
1
Open Claude's connectors
In Claude, go to Customize, then Connectors. Press + and choose Add custom connector. On the web it’s at claude.ai/customize/connectors.
2
Paste the address
Name it Marsell, paste
https://getmarsell.com/mcp and press Add.3
Sign in and press Allow
Claude opens Marsell. Sign in if it asks, check what Claude may do, and press Allow.
What you allow
The page Marsell shows says which app is asking and where it sends you back (claude.ai or chatgpt.com), which account it gets (your own or your team’s) with its projects, and what it may do. Can act is on, as for a setup’s key: the app may then answer what you ask it to, make pictures and spend tokens. Switch it off to let the app only look. Press Allow or Decline. An app connected this way works as you, with your role in Marsell checked on every call, exactly like a key. In an organisation, only administrators can allow an app in.Disconnect
Each connected app is a row in Your own AI on Agents, under Connected apps and keys: its name, where it’s from, when it was connected, when it was last used, and whether it reads and acts or reads only. Press Disconnect and it stops working on its next call. You can also remove Marsell in Claude’s or ChatGPT’s own settings. Connecting the same app again replaces its earlier connection.Try asking
- “What needs me on Marsell this week?”
- “Read my brand and my market, then tell me what we should work on next.”
- “What’s scheduled to post next week?”
- “Which posts did best this month, and why?”
Make a key
1
Open Your own AI
In the app, open a project and go to Agents. The section Your own AI is on that page.
2
Name the key
Give it a name you’ll recognise, such as “Claude on my laptop”. Turn on Can act if the agent may do things for you: answer what you ask it to, make pictures and spend tokens. Without it, the key Reads only.
3
Or make a missions key
For the AI that sends Marsell’s messages on Reddit, X and LinkedIn from your missions profile, press Make a missions key under Missions for your AI instead. That key can only take and report missions and record its timer, nothing else. See Missions for your AI.
4
Make it and copy it
Click Make a key. The key starts with
mk_ and shows only once: copy it, then click I’ve copied it.Connect your agent
The server is athttps://getmarsell.com/mcp. Claude Code, Cursor, OpenCode and Gemini CLI sign in with no key. Codex, Manus and other clients take a key: put it where mk_… is.
- Claude Code
- Codex
- Cursor
- OpenCode
- Gemini CLI
- Manus
- Any MCP client
/mcp in Claude Code, pick marsell and choose Authenticate. Sign in to Marsell in the browser that opens and press Allow.If Claude Code says marsell already exists, run claude mcp remove marsell and add it again.To use a key instead:What your agent can do
When you have more than one project, the agent names the project it means; with one project it can leave it out.
Marsell lists its tools in two ways. Most agents get the tools they use most in full (
list_projects, card_read, answer_decision, work_list, homework_list, homework_get and homework_take) and every other tool by its name and one line, with tools_manual first in the list. Claude Code still gets every tool in full, and so does a key for missions only. Before your agent first calls a tool it holds by one line, it reads it with tools_manual: by names, or by group (voices, series, setup, scheduled_posts, homework or missions) for what Marsell expects of that work. A call whose arguments fit runs at once; one that doesn’t is refused with what was wrong and the tool’s full instructions, so the next call can fix it. What each tool does is the same either way.
Everything Marsell’s chat can do
Your agent also gets the tools Marsell’s own chat works with, under the same names and with the same descriptions. When the chat gains a tool, your agent has it too. Every key gets the ones that only look, for example:place_open,agents_read,agent_openandwaiting_read: any place, any agent and every card that waits for you;freezer_read: the cards in the freezer, each with who made it, why it froze, when it went in and when it leaves;situation_read,work_listandcrm_search: where the business stands (with what is blocked, its one fix and where), the work planned and made, and the people;listings_read: the directory listings waiting for the business, with the listing kit to fill each form from;ad_results_read: what each ad and campaign brought, as Marsell counted it, against what it spent, with the network’s own claim beside it and never added in, then every ad visit and result and whether it went back to Google. With it come the records of every real ad (what it showed, what it was bought for, its days, spend and the network’s numbers, and how it stands among earlier ads bought for the same thing);onlykeeps one network, kind or state, andidgives one ad whole, with every day’s numbers, its approvals and its settings. Where Marsell doesn’t count a business’s results yet, it says so and still gives the spend and the network’s numbers. See How clicks and sign-ups are counted.post_optionsandad_offer: for a picture Marsell made, where and when it could be posted, and whether and where it could run as an ad, with who sees it, the days and the spend.ad_offeralso runs your own finished picture, video or pictures as they are, when you ask: see below. Givead_offerandad_cardcountries(two-letter codes, such asDEandAT) only when you named them for that ad: it runs there instead of the countries saved under Where may the ads run?, up to 25 on Meta.
work_planandagent_mode_set: plan work, or move an agent’s switch;set_goal,set_typeandset_aside: the project’s goal, its type, and Not for us on a community or a rival;owner_facts_readandowner_fact_set: your own facts under What Marsell knows, and a fact added, changed or taken out, which every plan, post and page reads first;needs_readandneed_set: what each agent still needs from you, as the line on its card on Agents lists it, with how to give each one, and one given by its key (an address, a link, a name, a list), saved where that agent reads it. Anything to copy comes as its text, such as a sign-up form to put on your own site. A password, a key, an account to connect or anything that spends is refused: a secret is never given through a tool, and the answer says where in Marsell you give it. The applicant’s details only MarPly reads come back only as on file. Saving is free; where giving one leads an agent to paid work afterwards (one more subreddit read, Ben’s email written again), its own words inneeds_readsay so.needs_readneeds only a key that reads;listings_mark: report back on listings the agent sent, many in one call;freezer_take_out: put a frozen card back in What needs you, saying its age (a post comes back without its day). Throwing a card away and Melt all stay yours, in the freezer;agent_run_early: ask one of the plan’s agents for one extra run now, with a reason you read in the activity feed. It spends from that agent’s tokens for the month, and only while the agent is on and neither resting nor waiting;web_research: look something up on the web, which spends tokens;business_read_quoteandbusiness_read_start: read the business again from the web, as setup did. The price comes first, and it starts only on your yes;offering_addandrival_add: after setup, a product or service your site doesn’t list, in your own words, with its price and pictures (the same name again changes it), and a rival you name, with its kind (direct, adjacent, or what buyers do instead), kept on Market’s list as yours. Both are free;ad_cardandad_answer: on your yes to a picture’s ad, its one card in What needs you, which spends only once someone who may spend presses Start there; or your no, after which that picture isn’t offered again. Posting a picture waits for your press in Marsell’s chat, so your agent asks for it throughask_marsell.
ask_marsell, and anything there that spends or posts waits for your press in Marsell’s chat.
Run your own picture, video or carousel as an ad
When you ask, your agent runs media you already finished as an ad, exactly as it is:- It sends the files with
get_upload_link(kindfile). curl’s answer gives each file’s attachment id. A picture or video you sent in Marsell’s chat, on Telegram or Slack has an id too. - It calls
ad_offerwithattachments(1 to 10 ids) andasked: true. One file is one ad. With two or more it also saysas:one_ad: one ad with its tall 9:16 version, shown in Stories and Reels (two pictures or two videos);carousel: 2 to 10 pictures, in the order given.
- The offer says where it runs, who sees it, which days and the spend, and keeps nothing: a key that can only look can ask it too. On your yes,
ad_cardkeeps the files and makes the card. An agent that didn’t make the offer can name the sameattachmentsandasonad_card.
- Pictures: PNG, JPEG or WebP, up to 30 MB, at least 600 pixels wide (500 for 9:16), from 4:5 to 1.91:1 or 9:16.
- Videos: MP4, 1 second to 60 minutes, at least 250 pixels wide. A MOV is refused: export it as MP4.
- Carousel cards: all square or all 4:5.
film_id runs a film Marsell made the same way. Keeping the files is free, and nothing spends until someone who may spend presses Start on the card. To have a new ad picture made from a photo, ask for a picture instead.
Set up a business with your AI
Your agent can set up a new project from what it already knows, such as your site, your code, your docs and your files, instead of Marsell’s paid read.1
Open the setup
At Start, press Connect your AI, then Make the command. The command carries a new key, which shows only once. If you typed your link first, the setup carries that address, and
get_setup shows it to your agent.2
Ask your agent
Run the command in your terminal, or add it to your client’s settings, then give your agent the prompt the page shows.
3
Check what it wrote
The brand board fills in as it writes. Click Check it so far, or Check what it wrote once it has finished, change anything that’s wrong, and confirm.
4
Connect your accounts and start
On the last step, connect the social accounts Marsell should post from, then press Start Marsell. Both are yours: your agent is told never to confirm or start the setup for you, and to ask you to connect your accounts before you start.
Nothing your agent writes is used until you check it and confirm. Setting up this way spends no tokens. After you confirm, Marsell files it all where its own read would, finds the communities and plans the first week.
Voices, including your own
Your agent can find a voice for your films and voice-overs, or clone your own.Clone your own voice
1
Record yourself
One to two minutes of you talking naturally, alone, in a quiet room, with no music. More than three minutes doesn’t help. MP3 is best; WAV, M4A, OGG, WebM and FLAC work too, and several files are fine.
2
Tell your agent it's your voice
Your agent clones a voice only when you have said, in your own words, that it’s your own voice or that the speaker agreed to it.
3
Let it send the recordings
Your agent runs one command from your machine, for example
curl -sS -F file=@part-1.mp3 -F file=@part-2.m4a <link>. Every file in that one command makes one voice: up to 25 recordings and 50 MB together, and Marsell hears at most 10 minutes of them. The link works for 30 minutes, and nothing big passes through your chat.4
Use it
Pick it by name in Create’s voice choice. Your films and voice-overs speak in it.
Series
A series is a recurring piece with its own name, on a schedule, with each episode posted to your accounts. Your agent can set one up for you, including a video series in your own voice, made and posted every day from the day’s news on a topic (daily archaeology news, say). Each day Marsell searches the news on the topic, leaves out what the series already reported, picks the best new story and makes a Short about it that credits its sources. A day with no new story gets no episode.
Your agent quotes the month first and tells you the tokens and the dollars. It starts a series only after you say yes in your own words, and Marsell keeps those words with the series. The month you agree to covers every kind of video on the schedule together. The month your agent gives must be the month as it is quoted now: if the price has moved since you said yes, up or down, the series doesn’t start and your agent asks you again with the new figure. This holds for every kind of series: a picture series and a newsletter are quoted and started on their month too, and for a picture series or a newsletter that month is also the most Marsell spends on it in a calendar month by itself.
Each episode follows your choice for series episodes on Settings, under What Marsell does on its own: off, ask each time, go ahead up to a budget, or no limit. When you start a series, you can also let its own episodes go ahead without a card, up to the quoted budget for one episode. That stays on this series only and changes none of your other choices; series episodes switched off, or the agent paused, still stop it. Each episode is charged what it really costs.
Homework for your AI
Some work needs your own logins, files or inbox, and no app can do it for you. Marsell puts all of it on one list, whichever agent it came from: posts for accounts Marsell can’t post to (YouTube posts and polls among them), UpSell’s Reddit replies, your directory listings as one batch, the applications MarPly found that you said yes to (grants, free credits, programs, competitions), the missions, real photos from your own files, the end card, jobs Marsell’s plan writes, a Google Ads account set up with you, and any card you hand over with Give it to my AI. Your agent picks it up on its own schedule, does it in your name and reports back.homework_list and homework_get take session too, and ignore it, so an agent that passes it on every call is never refused.
Each job is done in one of four places, named in needs:
profile: 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 on Reddit, X or LinkedIn in your name goes from it, and nothing else is done from it.mailbox: your own mailbox, as you, or a mail connector of your agent’s that sends from it. An email counts as sent only once it has left your mailbox: an agent that can only make a draft leaves it and gives the job back.files: your own files, on your computer or through a files connector of your agent’s.accounts: your own browser, signed in to your account on the site the job names.
missions_next, mission_inbox, and homework_take for a job whose needs is profile take in_missions_profile. A call without true does nothing (nothing is handed out, held, counted or reported) and answers with what the missions profile is and the choice: call again with true from that profile, or leave the work for a wake that is in it. A job done at its time, such as a post on X, isn’t takeable before that time; one scheduled ahead on the site, such as a YouTube post, is. The prompt your agent’s timer runs only starts this loop: every rule is on the list or in each job’s steps, so a timer set with older words meets them too. On a timer your agent can’t ask you anything: where a step says to ask you or show you something first, it leaves that part and reports it.
Jobs go on the list while Give my AI homework is on in Agents, Your own AI, and an agent that can act is connected. A new business is asked about it in setup. Your agent only runs when something starts it, and Marsell never calls it. So in setup, with homework on, your agent asks you whether it should go on a pulling schedule: a timer of its own that wakes it, calls homework_list and does what is due. It suggests once each morning, and offers every few hours too, so posts with a set time go out close to it. Each wake uses your agent’s own app, not Marsell’s tokens; only a job that shows spends_tokens uses yours. Only on your yes does it set the timer, with its own app’s scheduler (a scheduled task, a routine, an automation or a cron job) and the prompt Marsell gives it (timer_prompt). If its timer lasts only while the conversation is open, or stops after some days, it tells you so and when it stops. It records your answer with set_preferences homework_schedule, and finish_setup refuses until it has, unless homework is off. If you say no, or its app can’t schedule, you can say “check Marsell’s homework” to it any time.
After setup, while homework is on and no timer is on file, or the one on file has stopped, homework_list tells your agent to ask you again. homework_schedule records the answer at any time, and a no stands: your agent isn’t asked again. Marsell can’t see the timer: it keeps what your agent says it set. Your own AI shows whether your agent wakes on its own, what it set and until when, and when it last checked. If jobs wait and it hasn’t checked for two days, a card in What needs you says so, and what timer it said it set. When its timer runs an older prompt, Your own AI says so and has a line to paste that updates that same timer.
Marsell also sees when your agent reads its list but leaves the work, most often because its timer runs where it can’t reach the place the work needs. When work for one place could be taken through several wakes over more than a day and none of it is reported, you get one card and one email covering every place where that holds, saying what waits and where it’s done. Got it quiets it for a week while the reads still count, and the card closes once that work is reported or stops waiting. A take ends nothing; a report does. Nothing is ever held back from your agent.
The first kind of job is a batch of posts for a channel Marsell can’t post to itself: YouTube posts and polls, and posts for an account that isn’t connected. Those posts reach the list after your yes on their own cards. Marsell’s plan can also give your agent whole jobs, such as the end screen and pinned comment after an upload, review replies on Tripadvisor or Yelp, directory forms and event listings. Each comes to you first on a card showing everything your agent will do, and goes on the list when you press Give it to my AI. Only you answer that card, in Marsell: your agent can’t answer it, and neither can Marsell’s chat. Replies and messages to send from your own accounts, missions, directory listings and applications to grants and programs you said yes to are on the same list, each saying where it is done. Each finished job shows in Your own AI with its links.
A third kind is photos. Marsell makes pictures of real places and things only from real photos of them. When it finds none of a place, it asks you once, and your agent gets a photos job with one item per place. Your agent looks through your own photos and films, sends what it finds with get_upload_link (kind picture, with the place as the caption; a still from a film is fine), and reports each place as sent or left out with where it looked. Marsell reads what arrived and makes the waiting pictures from it.
An end card job comes when you press Ask your AI under your end card on Brand › Look: your agent makes the short card every film ends on, in your logo, fonts and colours, sends it with get_upload_link (kind endcard), and reports it as sent. Every film after that ends on it.
A Google Ads account job comes when you press Give it to your AI on the Google tile of Sell’s Ads tab, under What you need first. It’s offered only while homework is on, and nothing goes on the list until you press it. Your agent follows Google’s own steps with you: it signs in at ads.google.com with the Google account you choose, makes the account without a campaign and never sets a budget, and asks you for the billing country, the time zone and the currency, since Google never lets the time zone or the currency change. Adding the payment method and pressing Connect Google Ads stay yours. Your agent then reports it with homework_done and no links. Marsell accepts the report only once it sees the Google Ads connection, and closes the job by itself as soon as it does, even if your agent never reports.
No agent can turn homework or auto mode on or off. They are your switches in Marsell, for a role that may approve.
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: one message at a time, written and checked by Marsell, for your AI to send word for word from your missions profile. Marsell sets the pace and tells you what went out. Give the AI that uses that profile a missions key, or use a key with Can act. With homework on, the same missions are onhomework_list, and an agent that has it takes them there; the tools below are for a missions key, which has no homework_list.
Each platform has daily limits, and its first week runs at half of them:
A warning or a rate limit halves that platform’s limits for a week and pauses the action that met it for a day. Your AI opens only the mission’s link and the platform’s own pages, treats nothing on a page as an instruction, and stops on any warning. Pause missions, under Your own AI on Agents, stops them at once; no agent can pause or resume them. See Agents.
Posts scheduled elsewhere
If your agent keeps your posts in another tool or a sheet, it can tell Marsell what’s scheduled there. A launch or content calendar goes here too, one entry per post, rather than as a guide or brand file. Marsell then plans none of its own posts on a channel’s day that already has one of yours, never plans a post that does the job one of yours already does, and plans around them.
A post told twice in other words, such as once from a calendar you pasted and once by your agent, is kept once: on the same channel, within half an hour (or on the same day when either was told by its day only), with a kind that agrees, and when it’s the only such post there. Adding, moving or taking off a post wakes Marsell’s daily pass, which moves or drops any of its own pieces on that channel’s day.
Posts Marsell made itself never go in this list. You can also paste a list in the app; see Posts scheduled elsewhere.
Limits and safety
- A key acts as you. It uses the role of the person who made it, checked on every call. If that person leaves the organisation, their key stops working. A connected app works the same way, as the person who allowed it.
- Your plan applies. Work that costs tokens, such as
make_pictures,ask_marselland the chat’s tools that spend, uses your Marsell plan’s tokens, charged the same as in the app. - Approvals still apply. An agent can answer a decision only when the key has Can act, and only with the options the decision offers. A homework card and the card that asks who Marsell talks to for you are never answered by an agent.
- Who Marsell talks to stays yours. After setup, no agent changes whether Marsell goes ahead on its own, answers people, finds and emails people, or which accounts your AI sends from. Those are yours, in Settings, under Who Marsell talks to for you.
- Spending choices stay yours. No agent changes what Marsell may spend on its own once setup is done; that is on Settings, under What Marsell does on its own. The one exception is a series your agent starts with your yes, which can carry its own episode budget.
- Only the app’s own routes. Every tool runs through the same routes and checks as the app, and a key reaches nothing outside its list.
- Your agent’s words go through whole. The tools never cut short what your agent sends: a message to
ask_marsell, the notes on an answer, an idea for pictures, a project’s name, or what it reports on homework. - Rate limit. Up to 120 calls a minute per key or connected app.
Technical details
POST https://getmarsell.com/mcp, Streamable HTTP. Each request carries one JSON-RPC message, or a batch of up to 20, and gets JSON back. There is no event stream or session:GET /mcpanswers 405.- Methods:
initialize,ping,tools/listandtools/call. Protocol versions 2025-11-25, 2025-06-18, 2025-03-26 and 2024-11-05. - The tool list can be read without signing in, so you can see what Marsell does before you connect. Every tool call signs in: without a valid key or token it gets 401 with
WWW-Authenticate: Bearer resource_metadata="https://getmarsell.com/.well-known/oauth-protected-resource/mcp". Apps that sign in, such as Claude, Cursor and OpenCode, get that answer when they first connect, so they ask you to sign in straight away. - A missions key (made with Make a missions key) reaches only the six mission tools,
homework_scheduleand the list of projects. Any other tool answers that it’s outside what a missions key may do. A key with Can act reaches the mission tools too. Pausing and resuming missions are never reachable by any key.
Signing in (OAuth)
Claude, ChatGPT, Claude Code, Cursor and OpenCode sign in through Marsell’s own authorization server, following the MCP authorization spec (revision 2026-07-28):- Protected resource metadata:
/.well-known/oauth-protected-resource/mcp(and at the root). The resource ishttps://getmarsell.com/mcp; the authorization server ishttps://getmarsell.com. - Authorization server metadata:
/.well-known/oauth-authorization-server. - Dynamic client registration:
POST /oauth/register. Redirect URIs must be https, or http on127.0.0.1orlocalhost; they are matched exactly, loopback ones on any port. Public clients (none),client_secret_postandclient_secret_basicare supported. GET /oauth/authorize: authorization code with PKCE,S256only. Every answer carriesiss. Theresourceparameter must namehttps://getmarsell.com/mcp.POST /oauth/token:authorization_codeandrefresh_token. Codes last two minutes and work once; access tokens last an hour and are bound to/mcp; refresh tokens rotate on every use.POST /oauth/revoke(RFC 7009). Revoking a refresh token ends the connection.- Scopes:
read, andact, which includes read.
Not the same as the docs’ search server
https://docs.getmarsell.com/mcp searches these docs and needs no key. It can’t see your projects. See Use these docs with AI.