An agency-grade AI agent platform is one where the thing you resell is a client environment, not a seat. Three properties decide whether that works commercially: each client's agent is isolated from the others, the client sees a surface that carries their name, and the cost per client leaves room for your margin. Everything else on a feature list is preference.
The short answer. If you resell agents to small clients and want each one on its own domain with its own credentials, Sokko's Bakery plan is the cheapest way to get four isolated client agents. If you resell conversational support at volume, Botpress prices per conversation and is easier to quote. If the client insists on owning the stack, Dify self-hosted per client. If you are building a product rather than a service, LangGraph Cloud. If your clients' problem is reaching obscure SaaS tools, Zapier.
How these are ordered. By how much of the multi-client delivery problem each one removes for a two-to-ten person agency. Sokko is first, and Sokko writes this page, so treat the ordering as a bias and the table as the useful part. Prices were read from each vendor's own pricing page in November 2026.
Table of Contents
What an agency actually buys
A single agent for one client is easy and almost any tool can host it. The problem shows up at client four or five, when you are maintaining separate free-tier accounts, remembering which password belongs to which dashboard, and discovering that a client's security review rejects a shared runtime.
Three things have to be true before that point:
Isolation. Each client's agent runs somewhere the other clients' agents cannot reach. Not a folder, not a shared database with a
client_idcolumn, a separate process with separate credentials.A branded surface. The client's own domain, their login, their access list. This is what turns a tool subscription into a service line.
A price per client environment rather than per seat or per conversation, so you can quote a fixed retainer without modelling usage you cannot predict.
Most platforms on this list do one of the three well. The question is which one your delivery model depends on.
The seven platforms at a glance
| Platform | Tenant model | Branded surface | Price per client environment (November 2026) | Agency type it fits |
|---|---|---|---|---|
| Sokko | One isolated machine per agent | Custom domain, roles and SSO on Bakery; EU residency and EU-hosted inference | Bakery $99 per month for 4 agents, plus $25 per month for shared memory | Agencies running 2 to 4 client agents that must stay online |
| Botpress | Managed bot per client | White-label chat widget and webchat, custom domain | Plus $150 per month with 250 conversations; Team $750 with 1,500; overage $0.65 and $0.50 per conversation | Agencies reselling customer support chat at volume |
| Relevance AI | Managed multi-agent workforce | Client-facing branding on the built surface | Credit and action based; tiers are not published | Agencies selling GTM or ops outcomes, not hosting |
| Dify | Workspace per client, or self-hosted | Full control when self-hosted | Professional $59, Team $159 per workspace per month on annual billing | Agencies whose clients demand ownership of the stack |
| n8n | One instance per client | Whatever you build in front of it | Hosted from EUR20 per month; self-hosted free | Automation agencies whose work is workflows, not resident agents |
| LangGraph Cloud | Deployment per client, managed | Your own front end | Plus $39 per seat per month, plus resource time for deployments | Agencies with engineers building bespoke agent logic |
| Zapier Agents | Account per client | Client's existing Zapier surface | Professional $19.99, next tier $69 per month | Agencies plugging into a client's existing SaaS estate |
Read the table by unit. Per agent, per conversation, per workspace, per seat and per task are five different ways to lose money at scale, depending on what your client does all day.
1. Sokko
Sokko treats an agent as something that runs continuously rather than a chat window you open. Each agent gets its own isolated machine, which is the property an agency needs most: client A's agent cannot see client B's agent even when both are on the same plan.
For agency delivery the relevant tier is Bakery at $99 per month. It carries four agents, roles for your team, SSO, custom domains, 60 GB of storage, five parallel devboxes and 90 days of log retention, with $20 per month of included AI credits. At full occupancy that is roughly $25 per client agent per month, which is the number to put in a retainer model.
The rest of the platform matters for handover. Agents connect to Slack, GitHub, Notion, Gmail, Discord, Linear and calendars among 1,500 available integrations, and bring-your-own-keys support means model spend can sit on your account or the client's, whichever the contract says. EU data residency and EU-hosted inference are available, which decides a fair share of European deals before the technical questions start. Each agent is reachable at its own subdomain, and on Bakery it can carry the client's own domain.
Where it loses. Four agents on Bakery is a hard ceiling, so a fifteen-client roster means a conversation rather than a checkout. Custom domains and SSO only appear on the top plan. There is no per-client billing: you invoice the client yourself and reconcile by hand. And per-agent isolation costs more per unit than packing many clients into one shared runtime, which is the trade an agency makes deliberately.
2. Botpress
Botpress is the default choice when what you resell is a chat experience. Billing is per conversation, the widget is white-labelled, and clients recognise the shape of the product because it looks like the support chat they already use.
What it costs. Plus is $150 per month with 250 conversations included, Team is $750 per month with 1,500, and overage runs at $0.65 and $0.50 per extra conversation sold in packs of 100.
That unit is either an advantage or a trap. A support bot answering 200 short questions a month costs you a flat $150 and the margin is easy to explain. A client whose customers write long, multi-turn threads burns conversations faster than the retainer expects, and you meet that in the invoice rather than the contract.
Where it loses for agencies. Conversation pricing rewards high volume, short interactions. It is a poor fit for agents doing back-office work where there is no conversation to count: reading a mailbox, watching a repo, filing tickets on a schedule.
3. Relevance AI
Relevance AI sells a workforce of agents rather than infrastructure, and it prices in credits and actions. The public pricing page does not publish tiers, which means you cannot quote a client from the website. You model the workload first.
That is fine for agencies selling outcomes, because the conversation with the client is about the deliverable, not about compute. It is awkward for agencies selling hosting, where the client wants a fixed monthly number and you want to know your floor.
Where it loses. No published per-client price, no meaningful isolation story you can put in front of a client's security team, and a credit model that makes a fixed retainer a forecast rather than a fact.
4. Dify
Dify is the answer when the client's IT department has to own the thing. Workspaces are $59 per month on Professional and $159 on Team with annual billing ($590 and $1,590 a year), and the open-source edition can be installed on infrastructure the client controls.
For an agency that means two different businesses. Selling Dify Cloud per client is cheap and fast. Selling self-hosted Dify per client turns you into the operations team for every install: upgrades, backups, TLS, and a support call every time a plugin updates.
Where it loses. The workspace model is built for building RAG applications from a UI, not for keeping a resident agent alive with persistent state. Once a client's agent needs to remember yesterday, you are building the missing layer yourself.
5. n8n
Most automation agencies already run n8n, and for workflow delivery it is the best value on this list: hosted from EUR20 per month, free when self-hosted, and one instance per client is easy to reason about.
The limits appear when the client's request is not a workflow. Workflows trigger, run and stop. An agent that watches a shared inbox, keeps context across days, and decides what to do next is a different shape of software, and n8n makes you fake it with schedules and a database.
Where it loses. No resident agent, no persistent state you can hand to a client, and a self-hosted instance per client that you now patch.
6. LangGraph Cloud
LangGraph Cloud is what you reach for when the agent's logic is the product and the client's requirements do not fit a template. Plus is $39 per seat per month, with deployments billed on resource time on top.
For an agency this is the highest-margin and the highest-risk option at the same time. You can charge for genuine engineering. You also cannot deliver it without engineers, and every new client is a project rather than a configuration.
Where it loses. Time to a running agent is measured in weeks. Maintenance never ends because the code is yours. Agencies with one engineer and eight clients should not start here.
7. Zapier Agents
Zapier's advantage is reach. 9,000+ app integrations means the client's existing tools already connect, and at $19.99 or $69 per month the entry price is not the objection.
Agents inside Zapier are trigger-driven. They run when something happens, then stop. There is no persistent disk, so state lives in whatever app you connected, and an agent that needs to hold context across a week of work will not fit.
Where it loses. Always-on is not the design centre. If the client's problem is "reach this obscure SaaS tool", Zapier is the right answer and the other six are wrong.
The margin maths
Hosting is a small line in a retainer, and the platform choice moves it by an order of magnitude. Assume a client pays $750 per month for a managed agent. That is not a market rate, it is an arithmetic device, and you should use your own number.
| Platform | Cost per client per month | Share of a $750 retainer |
|---|---|---|
| Sokko Bakery at full occupancy (4 agents) | $24.75 | 3.3% |
| n8n hosted per client | ~EUR20 | ~3% |
| LangGraph Plus, one seat | $39 plus runtime | 5% plus usage |
| Dify Professional per workspace | $59 | 7.9% |
| Zapier Professional per client | $19.99 | 2.7% |
| Botpress Plus, 250 conversations | $150 | 20% |
The surprise is usually Botpress, not because it is expensive but because the unit is a conversation and conversations are what a support client buys most of. The same table also explains why agencies gravitate to per-agent hosting: the cost is fixed, so the margin is fixed, so the retainer can be sold with confidence.
Model spend is separate on every row. Bring-your-own-keys moves it to whichever party holds the API keys, and an agent making a model call every minute can cost more than the hosting line in a week.
Six questions to ask before you sign
Is each client's agent isolated, or are they all on one machine? Ask for the answer in writing, then ask what happens when one agent runs out of memory.
Can each agent reach the client's systems with its own credentials? Shared integration accounts become a breach story that names your agency.
What happens to the price when one client's agent runs all day? Fixed per agent is predictable. Per conversation and per credit are not.
Can I put the client's domain on it, and who manages the certificate? CNAME plus managed TLS is the baseline. Anything involving an afternoon of DNS work per client does not scale.
When the client leaves, can I get the logs and the data out? This is the question nobody asks during the sale and everybody asks at the end.
Is inference in the region my client's legal team will accept? Data residency is not the same thing as where the prompt is processed. The EU requirements guide covers why that distinction is worth checking.
Where Sokko is the wrong choice
More than four clients on one plan. Bakery stops at four agents. A twenty-client roster needs a commercial conversation, and if you need self-serve, look at Dify workspaces or n8n instances.
Reselling chat widgets at volume. Botpress and Flowise price per conversation or per prediction and are built for exactly that business.
A client who requires their own tenancy. Self-hosted Dify on their infrastructure answers the question their security team is actually asking, even though you inherit the maintenance.
One client, one agent. Almost any tool works. Paying for a fleet platform to run a single agent is overhead.
If you are still choosing a platform rather than planning client delivery, the broader comparison of ten AI agent platforms covers hosting models and state in more depth. If your clients are small businesses rather than enterprises, the small business shortlist is the closer read. And the white-label delivery guide covers the branding and tenancy work that sits on top of whichever platform you pick.