A chatbot answers; an agent does. It reads the ticket, looks up the order, drafts the refund and, if allowed, issues it. That last clause is why buying an agent platform is harder than buying a chat assistant: the software is being handed credentials to other systems and a budget it spends one model call at a time. This ranking covers thirteen platforms a small or mid-sized business can use to build and run such agents, from visual builders and automation suites to code frameworks with a hosted service and the agent products of three large vendors. We read each vendor's own pricing, documentation and legal pages and judged what a buyer can verify before signing: what the meter counts, which models can sit underneath and whose key pays for them, whether a human can be put in front of a risky action, where the thing runs, and what the vendor says it does with the content that passes through.
Visibility in this ranking can be paid for. Payment moves a vendor's position within the
shortlist; it never adds a vendor, and it never changes a word of the review. The largest vendors in AI agent platforms
cannot hold places 1 to 3. How it works: placement disclosure ·
editorial process.
How we ranked these
Setup: the first agent is easy, the first trusted agent is not
Most products here will produce a demo agent in an hour. The work that follows is what separates them: storing credentials for each connected system, deciding which tools the agent may call without asking, and writing the instructions that keep it on task. We looked at who can realistically do that work. Lindy and Dust are configured in conversation; Make, n8n, Dify, Flowise and StackAI use a canvas that rewards someone who thinks in steps; LangSmith assumes a developer. We also checked whether approval gates exist and where they work. n8n can pause an agent before a named tool and ask for a yes over Slack or email; StackAI's equivalent node functions only inside its chat interface. Self-hosted editions add a server to maintain, which is a setup cost that never ends.
The real price: thirteen products, almost as many meters
No two vendors count the same thing. n8n counts whole workflow runs, Make counts each module action, Relevance AI counts tool runs and model cost separately, Flowise counts predictions, Microsoft counts Copilot Credits at different rates per feature, Salesforce offers credits per action or a flat $2 per conversation, and Lindy, Dust and Langdock sell seats with a credit or usage allowance. An agent is a loop, so a per-step meter grows with how much the agent thinks, while a per-run meter does not. The second question is who pays the model provider. Dust includes the models in the subscription; n8n expects the buyer's own key; Make and Relevance AI allow either. We ranked higher the vendors whose bill can be estimated from the pricing page without a sales call.
Getting your data out: agent definitions, run logs and what happens to your prompts
Two separate questions sit under this heading. The first is portability. An agent is made of instructions, tool definitions and credentials, and no standard format carries all three between vendors, so a move means rebuilding. What can be kept are the prompts, which are plain text, and the run history, where the vendor exposes it. Self-hosted editions from n8n, Dify and Flowise keep definitions and logs in a database the buyer controls. The second question is what the vendor and its model providers do with content in transit. n8n, Langdock, Dust, Lindy, StackAI and LangChain state in writing that customer data does not train models, and Dust and Salesforce add that third-party providers retain nothing. Where we could not find such a sentence, the entry says so.
Independence from the vendor: two layers of dependence, not one
An agent platform sits on a model it usually does not make, so the buyer depends on the platform vendor and, through it, on a model provider whose prices, behavior and retirement dates the platform does not control. The defense is choice. Platforms that accept several providers and the customer's own keys, as n8n, Make, Langdock, Relevance AI and Flowise do, let a buyer change models without changing tools. Flowise and StackAI document local models, which removes the provider entirely for teams with the hardware. On the platform layer, a self-hostable edition is the strongest exit: n8n, Dify and Flowise can keep running on the buyer's servers whatever happens to the vendor's cloud. Agentforce and Copilot Studio are the opposite case, useful in proportion to how much of the business already lives with that vendor.
Who it's for: the builder matters more than the company size
The useful dividing line is who will build and own the agents. If it is a developer, a code framework with a deployment service, LangSmith with LangGraph, or a self-hosted builder such as Flowise, gives the most control for the least money. If it is an operations person who already maintains automations, n8n, Make and Dify extend a skill they have. If the agents are for knowledge workers asking questions and delegating chores, Dust, Langdock and Lindy meet them in chat. Regulated organizations that need on-premises deployment narrow quickly to StackAI, IBM watsonx Orchestrate and the enterprise editions of Dify and n8n. Companies committed to Microsoft 365 or Salesforce should price the incumbent's product first, because the data and the sign-on are already there.
The 13 tools, reviewed
#1 n8n
Workflow automation with agent nodes, cloud or self-hosted · Germany · n8n.io
Charged per run, not per step Free self-hosted edition Approval gate on agent tools
The meter is what puts n8n first. An execution is one run of a whole workflow however many steps it has, so an agent that loops through a dozen tool calls costs the same as one that makes a single call, and every listed plan allows unlimited users. The vendor says any LLM can be used, connected with an API key, and its docs describe an approval step that pauses an agent before a chosen tool runs, with the request delivered through Slack, Teams, Gmail, Outlook, Telegram and four other channels. The AI terms rule out training models on customer content.
Where it falls short
The free Community Edition leaves out single sign-on, projects, workflow and credential sharing, Git version control and log streaming, and on the cloud side SSO with SAML and LDAP sits on Enterprise only. The Business plan, at 667€ a month, is sold for self-hosted installs only. The pages we read describe connecting a model provider with an API key, so expect the model bill to arrive separately from the execution plan.
Wrong for
An office team with nobody who is comfortable reading a workflow canvas, wiring credentials and debugging a failed run. A chat-first product such as Lindy or Dust asks far less of the person building the agent.
Pricing: Cloud Starter is listed at 20€ a month billed annually for 2,500 workflow executions and Pro at 50€ a month for 10,000; the self-hosted Community Edition is free, and a self-hosted Business plan is 667€ a month for 40,000 executions. (free plan)
Visit n8n →
#2 Dify
Agentic workflow and RAG builder with a self-host edition · United States · dify.ai
Priced per workspace, not per seat Community edition on Docker Model providers via marketplace
Dify charges the workspace, not the person: Professional covers three members and 50 apps for one annual fee, Team covers 50 members and 200 apps. The free Sandbox is enough to build a working prototype, and the same product can later be run on a buyer's own machine, where the docs ask for only two CPU cores and 4 GiB of memory. Model providers are installed from a marketplace, and the pricing page names OpenAI, Anthropic, Gemini, xAI and Tongyi. Workflows can stop to ask a person for input before carrying on.
Where it falls short
The vendor describes the Community edition as source-available under an Apache-2.0-derivative license, which is not the same promise as an unmodified open source license, so read the terms before embedding it in a product. SSO, RBAC and audit logs are listed under Enterprise. The Sandbox allows one member and 200 message credits. The terms we read do not contain an explicit statement on model training either way.
Wrong for
A department that wants finished agents wired into its inbox and calendar on day one. Dify is a builder for applications and workflows; someone has to design, test and publish each one.
Pricing: The cloud Sandbox is free with 200 message credits; Professional is listed at $590 per workspace per year and Team at $1,590 per workspace per year; the Community edition is free to self-deploy and Enterprise is quoted. (free plan)
Visit Dify →
#3 Make
Visual automation platform with built-in AI agents · United States · make.com
Agents on every plan Own LLM key accepted Free tier with no time limit
Make is the least intimidating route from "we already automate things" to "an agent decides which automation to run". AI agents are available on all plans, and the pricing page says they work with Make's own AI Provider or with the customer's LLM key; the agents page lists OpenAI, Anthropic Claude, Mistral AI, Google Vertex AI and Azure OpenAI among the connectable providers. Rules, manual approvals and stop points can be set on an agent, and it reaches the same catalog of more than 3,000 apps the scenario builder does. The Free plan has no time limit.
Where it falls short
Every module action in a scenario spends a credit, so an agent that calls tools freely burns through an allowance faster than a fixed scenario would, and the Free plan caps at two active scenarios with a 15-minute minimum interval. Nothing we read offers a self-hosted edition. The privacy notice we fetched makes a no-training statement only about Google Workspace API data, not about customer content in general.
Wrong for
Anyone who must keep workflow data on infrastructure they operate, or who needs a written no-training commitment covering all content before signing. n8n and Langdock both put that in plain words.
Pricing: A Free plan includes up to 1,000 credits a month; Core is listed at $9 a month, Pro at $16 and Teams at $29, each starting at 10,000 credits a month; Enterprise is custom. (free plan)
Visit Make →
#4 Dust
Company-wide agents on your internal knowledge · France · dust.tt
Model choice included in the price EU or US data residency Zero retention at providers
Dust sells one subscription that already contains the model bill. The plans include GPT, Claude, Gemini, Mistral and DeepSeek models, more than twenty in all, so nobody has to open five provider accounts to compare them. The privacy policy names its foundational model providers, says they are prohibited from training on customer data, and says they operate under zero data retention. US or EU residency is offered on both the Business and Enterprise tiers, and SSO and a SOC 2 Type II report come with Business instead of being held back for the top tier.
Where it falls short
The Business plan allows up to three data-source connectors and five remote MCP servers; unlimited connectors, SCIM provisioning, audit logs and pooled credits are Enterprise items. The free allowance is 500 credits for the life of the account, not per month, so it is a tasting menu. We saw no self-hosted edition offered and no bring-your-own-key option on the pages we read.
Wrong for
A team whose agents mostly need to push buttons in other systems on a schedule. Dust is strongest as a question-answering and drafting layer over internal knowledge; n8n or Make are built around triggers and actions.
Pricing: A Free tier carries 500 lifetime credits; Pro is listed at 24€ a month on yearly billing or 30€ monthly with 8,000 credits per seat per month; Max is 120€ or 150€ with 40,000 credits; Enterprise is quoted. (free plan)
Visit Dust →
#5 Langdock
EU-hosted chat, agents and workflows for companies · Germany · langdock.com
Hosted in the EU Bring your own model keys SSO on the Business plan
Langdock answers the questions a European compliance officer asks before the demo ends. The application and most models are hosted in the EU on Azure, the security page says customer data is never used to train or improve AI models, and a company can plug in its own model provider keys. The company holds ISO 27001 and SOC 2 Type II. SSO, SCIM and SAML are part of the Business plan, which is unusual at €25 a seat. Agents can be shared with colleagues, used from Slack and Teams, and given actions through native connectors or MCP with confirmation steps.
Where it falls short
Serious workflow volume is a separate purchase: the base subscription includes 2,500 workflow runs a month, and the next step is an add-on starting at €539 a month per workspace. A governance add-on is free until January 1, 2027 and €3.50 per user per month after that. Dedicated deployment in the customer's own cloud is reserved for Enterprise contracts of 5,000 seats or more. The trial lasts seven days.
Wrong for
A five-person firm that wants an agent to run unattended back-office automations all day. The per-user price buys chat and agents for people; heavy unattended runs push the bill toward the Workflows add-on, where n8n is cheaper.
Pricing: Business is listed at €25 per user per month and Business Max at €99; a Workflows add-on runs from €539 a month for 40,000 runs to €1,199 for 100,000 per workspace; Enterprise is custom. (free trial)
Visit Langdock →
#6 Relevance AI
Multi-agent workforce builder for business teams · Australia · relevanceai.com
Two meters: actions and model cost Own API keys on every plan Pick AU, US or EU storage
Relevance AI is the rare vendor that itemizes. An action is one run of a tool, whatever happens inside it, and vendor credits are the pass-through cost of the model and any paid tools; the docs say bringing your own API keys bypasses vendor credits on every plan. That makes it possible to see whether an agent is expensive because it is busy or because it is using a costly model. Approvals and escalation paths are part of the workforce feature, and the customer chooses at signup whether data is stored in Australia, the US or the EU/UK.
Where it falls short
The docs state that the free plan is retired for new signups and existing accounts, so there is no way to build anything without paying. Pro has two build seats and no end users. Overage is steep: $80 per 1,000 extra actions. The security docs rule out training on customer data unless there is a specific partnership agreement, while the terms grant a license to use client data to improve the solution, and those two sentences deserve a question to sales. SSO and RBAC are Enterprise only.
Wrong for
A buyer who wants to try before paying, or who needs SSO below an enterprise contract. Dify and Make both have working free tiers, and Langdock includes SSO in its standard plan.
Pricing: Pro is listed at $19 a month on annual billing or $29 monthly with 2,500 actions and $20 of vendor credits; Team is $234 annual or $349 monthly with 7,000 actions; Enterprise is custom. The free plan has been retired. (none)
Visit Relevance AI →
#7 Lindy
AI teammate that works through Slack, iMessage and email · United States · lindy.ai
Works from Slack and iMessage Approval before writes and sends Every action logged
Lindy skips the canvas. The agent is addressed like a colleague in Slack, iMessage or email, and the site promises that nothing irreversible happens without approval: actions that write or send can wait for a yes. Every action is logged with what happened and when. The terms say contributions are not used to train AI models and that the model providers have agreed to the same, and the company reports a SOC 2 Type II audit. Credits are only spent while the agent is working, so an idle month costs the seat fee and nothing more.
Where it falls short
No permanent free tier appears on the pricing page; the only free period described is a first week for teammates who join through Slack. Credits are the opaque part, since the page does not say how many a typical task consumes. Audit logs and a signed BAA are Enterprise features. The only postal address in the terms is a Beaverton, Oregon mailing address that a second vendor on this list also uses, under California governing law.
Wrong for
A process owner who needs a repeatable, versioned workflow with branching logic that an auditor can read. Lindy is an assistant you delegate to; n8n or Flowise give a diagram of every step.
Pricing: Plus is listed at $29.99 a month per user with 3,000 credits, Pro at $99.99 with 15,000 and Max at $199.99 with 35,000; Enterprise is quoted. (free trial)
Visit Lindy →
#8 Flowise
Open source visual builder for agents and LLM flows · United States · flowiseai.com
Runs locally with npm Local models via Ollama Flat hosted tiers
Flowise is the quickest route to an agent running entirely on a machine the buyer owns. The site describes it as open source, the install is a single npm command, and the docs list chat model nodes for OpenAI, Anthropic, Google, Mistral, AWS Bedrock and Azure OpenAI next to Ollama and LocalAI for models that never leave the building. Three builders cover different skill levels, from Assistant for a first agent to Agentflow for multi-agent systems, and a human-in-the-loop feature lets a person review what an agent did. The hosted plans bill by prediction at a flat monthly fee.
Where it falls short
The terms allow FlowiseAI to use anonymized and aggregated usage data to improve the platform and do not address model training on customer content in so many words. The hosted Free plan stops at two flows and 100 predictions a month. The terms give a Beaverton, Oregon mailing address for the Delaware corporation, the same one another vendor here lists, so the site tells a buyer little about where the team sits.
Wrong for
A company without anyone to install, update and back up a Node application, and without a developer to build each flow. Make and Lindy are aimed at people who will never open a terminal.
Pricing: Self-hosting the open source build is free; the hosted Free plan allows 2 flows and 100 predictions a month, Starter is $35 a month for 10,000 predictions and Pro is $65 a month for 50,000. (open source)
Visit Flowise →
#9 LangSmith
Deployment and observability for code-built agents · United States · langchain.com
MIT-licensed LangGraph underneath Deploys agents from any framework Self-hosting on Enterprise only
LangSmith is what a software team buys after its prototype works. LangGraph, the orchestration library, is MIT-licensed and free, supports human-in-the-loop checks and any model provider, and carries no obligation to pay LangChain anything. The paid platform adds tracing, evaluation and managed deployment, and the vendor says agents built with any framework can be deployed on it. A no-code product called Fleet lets non-engineers build agents with approval on sensitive actions and a choice of OpenAI, Anthropic, Gemini or custom models. The terms commit to not training on customer data.
Where it falls short
Three meters run at once, seats, traces and compute units, and only the first is predictable. Self-hosted and hybrid deployment, custom SSO and fine-grained RBAC are Enterprise items that need a sales conversation. The Developer plan is capped at one seat. The terms name a Delaware corporation and no street address; the about page gives San Francisco as headquarters.
Wrong for
A business with no engineers. The core of this product is a library and an operations console for people who write Python or TypeScript; Make, Dust or Lindy get a non-developer to a working agent far sooner.
Pricing: Developer is $0 for one seat with up to 5,000 base traces a month; Plus is $39 per seat per month with up to 10,000; usage beyond that is pay-as-you-go, with compute at $1.50 per LangChain Compute Unit; Enterprise is custom. (free plan)
Visit LangSmith →
#10 Microsoft Copilot Studio
Low-code agent studio for the Microsoft 365 estate · United States · microsoft.com
Published credit rate table No charge for M365 Copilot users OpenAI and Anthropic models
For a company whose staff already sign in to Teams every morning, Copilot Studio removes the distribution problem: agents publish to Teams, Microsoft 365 Copilot and websites from one low-code studio. Microsoft documents the meter in unusual detail. A classic answer costs 1 Copilot Credit, a generative answer 2, an agent action 5, and grounding on the tenant graph 10, and employees with a Microsoft 365 Copilot license use agents at no charge. The model menu goes beyond OpenAI: Claude models are listed as generally available, and admins can allow external models from Anthropic, xAI or Mistral.
Where it falls short
The credit arithmetic is the weakness. One reply can stack several rates, reasoning models add a per-token premium, and when a tenant reaches 125% of prepaid capacity its custom agents are disabled until capacity is added. The docs flag many model and region combinations as cross-geo, meaning data may be processed outside the customer's region. The free trial lets a maker build and test an agent and does not allow publishing it.
Wrong for
A firm on Google Workspace or a mixed stack, which would be buying into the Microsoft admin centers mainly to run one agent. n8n, Make and Dust do not care which office suite sits underneath.
Pricing: Sold tenant-wide in packs of 25,000 Copilot Credits at $200 per pack per month, with a pay-as-you-go meter as the alternative; users holding a Microsoft 365 Copilot license at $30 per user per month, paid yearly, are not charged for agent use. (free trial)
Visit Microsoft Copilot Studio →
#11 StackAI
Enterprise agent builder with on-premise deployment · United States · stackai.com
On-premise and VPC deployment No-training clause in the terms Nothing between free and quote
StackAI is built for the buyer whose security team has a veto. The terms say the company does not train any models on customer data and does not share it with AI providers to improve their products, and the site lists HIPAA, GDPR, SOC 2 Type II and ISO 27001. Deployment ranges from multi-tenant cloud to air-gapped on-premise. The docs cover models hosted on Azure and AWS Bedrock and a local LLM option, plus governance controls over which models staff may use. A free tier with 500 runs a month is there to evaluate it.
Where it falls short
There is no self-serve paid plan: one seat and two projects on Free, then a sales conversation, which makes it hard to budget before talking to a rep. SSO, on-premise and VPC deployment are listed under Enterprise. The Human in the Loop node only works inside the Chat Assistant interface, not when a workflow is triggered by API or used in Form mode, which is a real limit for back-office automations.
Wrong for
A small company that wants to swipe a card and know the monthly cost. Dify, Make and Dust publish prices for the plan most small teams would end up on.
Pricing: A Free plan at $0 a month covers 500 runs a month, 2 projects and 1 seat; the only other plan is Enterprise, priced on request. (free plan)
Visit StackAI →
#12 Salesforce Agentforce
Agent platform built into the Salesforce CRM · United States · salesforce.com
Per action or per conversation Zero retention at third-party LLMs Tied to Salesforce data
Agentforce earns its place for one kind of buyer: the company whose accounts, cases and orders are already Salesforce records. The agent reasons over that data without an integration project, and the Trust Layer masks personal data before a prompt reaches a model and states that prompts and responses are never stored by, or used to train, the third-party LLMs. Prices are public, which is more than most enterprise vendors manage. A no-cost Salesforce Foundations offering includes the builder tools, and verified MCP servers from AgentExchange partners extend what an agent can reach.
Where it falls short
Three meters coexist: Flex Credits at 20 per standard action, $2 per conversation for customer-facing agents, and per-user add-ons from $125 a month, and the pricing page says conversations and Flex Credits cannot be combined in the same org. Editions that bundle credits start at $550 per user per month. The product page we read does not say which model providers sit behind the agent.
Wrong for
Any company that does not run its sales or service operation on Salesforce. The value is the data already in the platform; without it, this is an expensive way to reach tools that Make or n8n connect to for a fraction of the cost.
Pricing: Flex Credits cost $500 per 100,000, with a standard action using 20 credits; customer-facing agents can instead be billed at $2 per conversation; per-user add-ons start at $125 per user per month. (free plan)
Visit Salesforce Agentforce →
#13 IBM watsonx Orchestrate
Agent control plane for large, regulated organizations · United States · ibm.com
Runs on-premises or multi-cloud Published starting prices External models through a gateway
IBM sells this as one place to build and manage every agent a large organization runs, and the feature a regulated buyer notices first is deployment: a managed service on IBM Cloud or AWS, or an install in the company's own data center. Oversight tools cover access and policy enforcement, quality evaluation and token cost tracking. The docs describe an AI model gateway for registering external models next to the built-in ones. Unlike most enterprise agent products, the starting prices are printed, and the 30-day trial includes agent building and the tools catalog.
Where it falls short
Essentials starts at $530 a month and the next plan starts at $6,360, a twelvefold jump with nothing in between. The plans are described in messages and monthly active users, a meter that fits chat-style agents better than background automations. Data isolation and the HIPAA-ready option sit on the quoted Premium tier. The product and pricing pages we read make no statement about training on customer data.
Wrong for
A small business or a single department testing its first agent. The floor price and the enterprise framing assume a platform team; Dify or n8n cost a fraction and can be abandoned without regret.
Pricing: Essentials starts at $530 per month and Standard at $6,360 per month; Premium is quoted; a 30-day free trial is offered. (free trial)
Visit IBM watsonx Orchestrate →
What the data says about this market
Of the thirteen platforms ranked, nine are headquartered in the United States, two in Germany (n8n and Langdock, both in Berlin), one in France (Dust) and one in Australia (Relevance AI). Twelve publish a paid price; StackAI alone lists only a free tier and a quote. Eight offer something permanently free, either a hosted free tier or an edition the buyer can run at no charge, four offer only a time-limited trial, and Relevance AI states that its free plan has been retired. The spread of meters is the more telling number: across thirteen products we counted executions, message credits, module-action credits, per-seat credits, tool-run actions, predictions, traces and compute units, Copilot Credits, Flex Credits, conversations, runs and monthly active users.
Deployment choice is wider than in most software categories. Seven of the thirteen state some way to run the platform outside the vendor's shared cloud: n8n, Dify and Flowise through editions a customer installs, and LangSmith, StackAI, Langdock and IBM through enterprise contracts. On data use, eight vendors put a no-training commitment in a page we could read (n8n, Langdock, Dust, Lindy, LangChain, StackAI, Relevance AI with a stated exception, and Salesforce for its third-party models). For the other five we found no explicit sentence on the pages fetched, which is an absence of evidence, not proof of the opposite, and a question to put in writing before signing.
The category is young enough that its vendors come from four different directions: automation suites adding a reasoning step (n8n, Make), application builders for language models (Dify, Flowise, StackAI), workplace assistants that learned to act (Dust, Langdock, Lindy), and incumbents attaching agents to an installed base (Microsoft, Salesforce, IBM). It also rides a longer trend. ICT services rose from 9.1% of world service exports in 2013 to 14.48% in 2023 (World Bank), and an agent that files a ticket or reconciles an invoice is that same trade in services, delivered by software billed per run.
The 13 ranked vendors, counted
- Headquarters by region: North America 9, Europe 3, Asia-Pacific 1
- By country: United States 9, Germany 2, Australia 1, France 1
- Pricing model: Free + paid tiers 4, Usage-based 3, Per user / month 2, Flat monthly 1, Open source + paid tiers 1, Per seat / month 1, Quote only 1
- Free option: Free plan 7, Free trial 4, None 1, Open source 1
Counted from the 13 vendors on this page. More in our market data.
For the wider market behind AI agent platforms, read our report The Global Shift to ICT Services,
or browse all industry reports.
Questions and answers
What is the best AI agent platform in 2026?
n8n leads this ranking for small and mid-sized businesses. It charges per workflow run regardless of how many steps an agent takes, allows unlimited users, has a free self-hosted edition, can hold an agent at an approval step before a chosen tool runs, and states that it does not train models on customer content. Dify is the better pick for a team that wants a free hosted sandbox and workspace pricing without seat counts. Make suits people who already automate with a visual builder and want agents on every plan, including the free one.
What is the difference between an AI agent and a chatbot?
A chatbot produces text in reply to a message. An agent is given tools, such as the ability to search a CRM, send an email or update a record, and the model decides which to use and in what order to reach a goal. The practical difference for a buyer is risk and cost: an agent holds credentials to other systems and may make many model calls per task. Several products here do both, and Copilot Studio still bills a scripted "classic answer" at a lower rate than a generative one.
How do AI agent platforms charge?
By almost every unit imaginable. Among the thirteen ranked here, n8n bills per workflow execution, Make per module action, Flowise per prediction, Relevance AI per tool run plus model cost, Microsoft per Copilot Credit with a different rate for each feature, and Salesforce per Flex Credit or at $2 per conversation. Lindy, Dust and Langdock charge per user with a usage allowance. LangSmith combines seats, traces and compute units. Convert each to the cost of one real task before comparing.
Can I use my own OpenAI or Anthropic API key?
On several of these, yes. n8n tells builders to plug in their own key for any LLM. Make's pricing page says agents run on Make's AI Provider or the customer's own LLM key. Relevance AI documents bring-your-own keys on every plan as a way to bypass its vendor credits, and Langdock accepts a company's own model provider keys. Dust and Lindy bundle model usage into the subscription instead, and we did not see a key option on the pages we read for either.
Which AI agent platforms can be self-hosted?
Three have an edition a customer can install without a contract: n8n's Community Edition, Dify's Community edition, which runs on Docker, and Flowise, which installs with npm. LangSmith offers self-hosted and hybrid deployment on its Enterprise plan, StackAI lists on-premise and VPC deployment under Enterprise, IBM watsonx Orchestrate can run on a company's own servers, and Langdock reserves dedicated deployment for contracts of 5,000 seats or more. Self-hosting the platform does not by itself keep prompts in-house; that also depends on which model it calls.
Do these vendors train AI models on my company's data?
Eight of the thirteen say no in a document we could read: n8n, Langdock, Dust, Lindy, LangChain, StackAI, Salesforce (for its third-party LLMs) and Relevance AI, whose security docs add an exception for customers with a specific partnership agreement. For Dify, Make, Flowise, Microsoft Copilot Studio and IBM watsonx Orchestrate we did not find an explicit statement on the pages fetched. That is not evidence that they do train on it, only a reason to ask for the commitment in the contract or data processing agreement.
How do I stop an agent from doing something it should not?
Use the approval features and limit the tools. n8n can require a human to approve a specific tool call, with the request sent to Slack, Teams, email or a chat app. Lindy holds actions that write or send until approved. Relevance AI has approvals and escalation paths, Make lets a builder add manual approvals or stop points, and LangSmith Fleet gates sensitive actions. Check where the gate works: StackAI's Human in the Loop node functions only in its chat interface, not for runs triggered by API.
Is an open source agent framework enough, or do I need a platform?
A framework such as LangGraph, which LangChain describes as MIT-licensed and free, is enough to build an agent if a developer is available. What it does not provide on its own is the place to run the agent continuously, a record of every run, credential storage, approvals and a way for non-developers to make changes. Those are what the paid LangSmith platform, or a builder such as n8n, Dify or Flowise, adds. Teams without engineers should start from a platform.
Which platforms keep data in the European Union?
Langdock says its application and most models are hosted entirely in the EU on Azure. Dust offers EU or US hosting on both of its tiers. Relevance AI lets a customer choose storage in Australia, the US or the EU/UK at signup. The self-hostable editions of n8n, Dify and Flowise can run in any data center the buyer picks. Microsoft's documentation marks many Copilot Studio model and region combinations as cross-geo, meaning data may be processed outside the customer's region.
Should a Microsoft 365 or Salesforce customer just use the incumbent's agent product?
Price it first, then compare. Copilot Studio costs nothing extra for employees who already hold a Microsoft 365 Copilot license and publishes straight to Teams. Agentforce works on the records already in Salesforce. Those are real advantages. The drawbacks are the meters, Copilot Credits with stacked rates and Flex Credits at 20 per standard action, and the fact that agents built there are of little use outside that vendor's estate. A cross-tool agent is usually cheaper on n8n or Make.
Why are some well-known names missing from this list?
Every product fact on this page was checked on the vendor's own site, and a product was left out when a required fact could not be confirmed there. In this round we could not read a headquarters address on the public pages of Zapier, Gumloop, CrewAI or Activepieces, and the pricing pages of OpenAI and Google Cloud did not load for us. Their absence is a statement about what we could verify, not about those products, and they will be reconsidered at the next review.
