If you run support for other people's sites, you have probably already hit the wall. You sign three clients, the AI agent is great, then one client has a launch and burns through the month's messages in a week. Now the other two clients have a dead bot, and you are the one explaining why. We see the exact question in our inbox: "How do you set up multiple accounts so one client doesn't use all the credits?" and "For 5000 messages I can serve only 7 clients per day."
That math is real, and it is the reason a shared message pool is the wrong foundation for an agency. The answer is not a bigger pool. It is isolation. Each client should run on its own budget, so a busy client can never drain a quiet one, and your cost per client stays predictable no matter who goes viral.
Why one shared pool breaks the agency model
A single capped pool of AI messages has one failure mode, and it is the worst one: it is shared. Every client draws from the same bucket. So your worst month is set by your loudest client, not by your average. You cannot price a client reliably when their cost depends on what every other client did that month. And you spend your time rationing instead of selling.
The fix has two parts, and muro ships both today. First, you do not pay per project, so adding clients does not inflate your bill. Second, each client can run the AI on their own Claude key, at cost, with no shared budget to fight over. Let us take them in order.
Part one: unlimited projects, one login, one bill
Our Fleet plan is $59/mo for unlimited projects. Not five, not twenty. Unlimited. That is deliberate, because adding a site costs us almost nothing, so charging you per client would be a tax with no cost behind it. You add a client by adding a site, paste one line of widget code, and you are live.
One login can also belong to multiple workspaces and switch between them, which matters for the next part. So your account stays simple while each client gets a clean boundary.
Part two: a separate Claude key per client, at cost
This is the piece that actually solves the budget problem. With BYOK (bring your own key), you paste an Anthropic (Claude) API key and muro's AI runs on that account. It is unlimited and at cost. muro adds zero markup. You are paying Anthropic directly for the tokens that client's visitors actually used, and nothing on top.
Here is the honest detail, because it changes how you set this up. The AI key in muro is set per workspace, not per individual site. So to give each client a truly separate AI budget, you give that client their own workspace and paste their key there. A busy client spends on their key. A quiet client spends on theirs. Neither can touch the other, because they are different accounts at Anthropic. There is no shared pool to drain.
- 01Create a workspace for the client (your one login can hold many).
- 02Add their site(s) and drop in the widget snippet.
- 03In Settings, AI, paste that client's
sk-ant-…key. The credit meter switches off for that workspace. - 04Turn on the AI agent and point it at the client's docs URL so it answers grounded on their product.
- 05Repeat for the next client. Each one is isolated by construction.
Where your margin actually comes from
Someone in our inbox put it plainly: "I'm considering to start my own AI agency with this bot." Good. Here is the honest version of the economics, because we would rather you build on real numbers than a pitch.
Your two costs are muro's flat fee and the AI tokens. With BYOK, the AI tokens are billed at cost with no muro markup, so for the client who owns their key, your AI cost is literally zero and you charge for your service on top. For clients on your key, your raw AI cost is the tokens themselves, again with nothing added by muro, so your headroom is the gap between that token cost and your retainer. Either way, your downside is capped and your price to the client is something you control, not something a vendor's per-resolution meter controls.
- →Flat platform cost: $59/mo Fleet, unlimited clients. Your tenth client does not raise it.
- →AI cost: the client's tokens, at cost, zero muro markup. BYOK is unlimited.
- →No per-seat fee, so you and the client's staff can all reply without growing the bill.
- →No per-resolution AI charge, so a client's good month is not a scary invoice.
If you do want a managed option for smaller clients who will not set up their own Claude account, every plan includes a monthly pool of credits, and you can buy one-time top-up packs when a client runs hot, without changing the subscription. But for an agency, per-client BYOK is the setup that scales cleanly.
What the client gets, and what they do not
Each client's AI is grounded only on their own docs. Point the agent at a URL and it builds a help center from that page, then answers visitors from that product alone. It is instructed never to invent prices or features, and it hands off to a human on refunds, cancellations, account-specific actions, a frustrated visitor, or anything it is unsure about. The moment a human replies, the AI backs off for that conversation. It is always labelled as AI and never pretends to be a person.
A concrete five-client setup
Say you run support for five clients. One workspace each, five Claude keys, one Fleet subscription, one login. Client A launches and their visitors send 4,000 messages in a week. That spends on Client A's key. Clients B through E are untouched, still answering instantly, because there was never a shared bucket for A to empty. You did not ration anyone. You did not get a surprise invoice. And when you sign Client F next week, you add a workspace and a key, and your muro bill does not move.
That is the whole model: unlimited projects so growing your client list is free, and a separate Claude key per client so no one shares a budget. Honest about the boundary (isolation lives at the workspace level), honest about white-label (roadmap, not shipped), and honest about cost (at cost, zero markup). If you are building a support agency on top of an AI agent, start a free trial and see how Fleet pricing works.