On-premises AI agent platforms
On-premises AI agent platforms compared: 15 options by type (2026)
Which AI agent platforms can run on your own infrastructure?
Agent platforms whose vendors say they install on servers you control include DataRobot, Dify, IBM watsonx Orchestrate, Katonic AI, Kore.ai, Lyzr, Onyx and VDF AI. NVIDIA AI Enterprise and Red Hat OpenShift AI supply the layer underneath, n8n and UiPath add agents to process automation you host, and CrewAI, LangGraph and Microsoft Agent Framework are code libraries your engineers build on. They are not substitutes for each other, so pick the type first and the vendor second.
The shortlist
| If you need | Shortlist first | Type |
|---|---|---|
| One governed place to build, run and use agents on your own servers, where the vendor states it can run air-gapped | Dify Enterprise, Katonic AI, Lyzr, Onyx, VDF AI | Agent platform |
| An agent platform from a large supplier you already buy from, with an on-premises option | DataRobot, IBM watsonx Orchestrate, Kore.ai | Agent platform |
| The base layer that serves models on GPUs you own | NVIDIA AI Enterprise, Red Hat OpenShift AI | Infrastructure platform |
| Agents added to workflow or process automation you host yourself | n8n, UiPath | Automation platform |
| A code library your own engineers will build agents on | CrewAI, LangGraph, Microsoft Agent Framework | Developer framework |
Names in each row are in alphabetical order. This is not a ranking. Katonic AI publishes this page and is one of the fifteen, so it is described with the same fields as every other platform and from the same kind of source: what the vendor says in public.
Facts checked on 22 November 2026 against each vendor’s own pages. Every statement about a vendor below carries a number that links to the page it came from. Where a vendor’s pages did not state something, this guide says "not stated on the pages we read" and does not guess. Products change, so read the linked page before you rely on a line here.
What "on your own infrastructure" means here
Buyers use four phrases for nearly the same thing, and vendors do not use them consistently. It helps to fix the words before comparing products.
| Phrase | What it means in this guide |
|---|---|
| On-premises | The software runs in a data centre or server room you control |
| Self-hosted | You install and operate the software yourself, on your premises or in a cloud account you own |
| Private cloud or VPC | The software runs in a cloud account or virtual network that belongs to you, not to the vendor |
| Air-gapped | The system runs with no connection to any outside network, including for licence checks, updates and model calls |
A product can be self-hosted and still call an outside service at run time, for a model, a licence check or usage reporting. The test for air-gapped is to disconnect the network and see what stops. If your regulator or your security team asks for air-gapped operation, ask each vendor to state it in writing for the exact edition you are buying.
The criteria used
Every platform below is described against the same seven questions. They are the questions a bank, a ministry or a hospital is asked by its own risk and audit teams.
- Where it runs. Vendor cloud, your cloud account, your data centre, or fully disconnected. What it installs on: Kubernetes, Red Hat OpenShift, Docker or an appliance.
- Who operates it. Your staff, the vendor, or a mix. A product operated by the vendor inside your account is not the same as one your staff run.
- Governance and audit. Access by role, single sign-on, a record of every model call and agent action, guardrails, approval by a person, and testing before release.
- How agents are built. A visual builder for business teams, code for engineers, or both. Whether there are templates to start from.
- Model choice. Whether you can use open-weight models served on your own GPUs, and whether you can change model without rebuilding the agent.
- How it is priced. By user, by run, by GPU or by quote. Whether the price for the on-premises edition is published at all.
- What it does not do. Every product in this list leaves something to you or to another product. The limits are as useful as the features.
These seven map to the four questions of the Sovereignty Test: who governs the system, who runs it, who holds the keys and who can switch it off [59]. Where it runs and who operates it answer the first two. Model choice and offline operation answer the third. Licence terms, licence checks and the vendor’s ability to stop the service answer the fourth.
The four types
| Type | What you get | What you still need |
|---|---|---|
| Agent platform you run yourself | A product for building agents, connecting them to data and tools, governing them and giving staff a place to use them | Infrastructure to run it on, and models |
| Infrastructure platform | Software for serving models and running AI workloads on your GPUs | The layer where business teams build, govern and use agents |
| Automation platform with agent features | Workflow or robotic process automation, with agent steps added | A check that the agent features exist in the self-hosted edition |
| Developer framework | A code library for writing agents | Everything around the code: access control, audit, a user interface, hosting and operations |
Most organisations end up with one product from the infrastructure row and one from the agent platform row, and many engineering teams also use a framework inside the platform. The mistake to avoid is comparing a framework with a platform as if they were alternatives.
Agent platforms you run yourself
Eight products, in alphabetical order. Each is sold as a place to build and run agents and each has an edition the vendor says can be installed outside the vendor’s own cloud.
DataRobot
| Best for | Organisations that already run DataRobot for predictive models and want agents under the same governance |
|---|---|
| Where it runs | Software as a service, a virtual private cloud or on-premises. The vendor keeps a separate documentation set for its Self-Managed AI Platform. Air-gapped operation is not stated on the pages we read |
| Who operates it | DataRobot for the managed service. The self-managed edition is installed in your environment |
| Governance and audit | The vendor describes access control for agents and users, approval workflows for deployments and compliance documentation generated by the platform |
| How agents are built | Low-code or code-first, with application templates and a command line tool |
| Models | You choose the models and frameworks. The documentation covers an LLM gateway, provider fallback and NVIDIA NIM |
| How it is priced | Not published. A 30-day trial is offered |
| Limits to check | Ask what the self-managed edition installs on and whether it runs disconnected, since the pages we read do not say. The self-managed edition has its own version line, so ask how far it is behind the managed service |
Sources for DataRobot, read on 22 November 2026: home page [1], AI Platform [2], trial [3], self-managed documentation [4].
Dify
| Best for | Teams that want an open-source visual builder they can start with for free and move to a supported enterprise edition |
|---|---|
| Where it runs | Dify Cloud, a free Community Edition you host with Docker Compose, and Dify Enterprise for your own virtual private cloud, your premises or an air-gapped network, installed on Kubernetes with a Helm chart |
| Who operates it | You, for the Community Edition. Dify Enterprise can be run by Dify as a single-tenant cluster or by you |
| Governance and audit | Dify Enterprise lists single sign-on, access by role down to the workflow, audit logs that can be sent to a security monitoring system and traces for each call. The pricing page lists single sign-on as an Enterprise feature |
| How agents are built | A visual workflow studio, with an API, a command line tool and plugins for engineers |
| Models | You choose the models. The Enterprise page says it can run offline with locally deployed models |
| How it is priced | Community Edition is free. Dify Cloud is priced per workspace: Professional at $590 a year and Team at $1,590 a year. Enterprise is priced on request |
| Limits to check | The source code is under a modified Apache 2.0 licence. It does not allow running a multi-tenant service on the source code, or removing the Dify logo, without written permission from Dify. A provider that wants to sell agents to its own customers should read the licence first. Air-gapped deployment, single sign-on and audit logs are described for the Enterprise edition |
Sources for Dify, read on 22 November 2026: Dify Enterprise [5], self-hosting documentation [6], pricing [7], licence [8].
IBM watsonx Orchestrate
| Best for | Large organisations that already run IBM software and Red Hat OpenShift and want one supplier for agents and the layer below |
|---|---|
| Where it runs | IBM Cloud, AWS, or your own premises. The documentation has installation guides for Kubernetes and for IBM Software Hub. It also has a topic on using Python tools in an air-gapped environment. Confirm with IBM what a fully disconnected install covers |
| Who operates it | IBM for the managed service. On-premises responsibilities are not stated on the pages we read |
| Governance and audit | IBM describes consistent access, policy and guardrails across agents, tools and workflows, with evaluation and observability. The documentation lists a governance assessment for each agent version |
| How agents are built | A visual agent builder and workflow builder, a development kit for code, a catalogue of prebuilt agents, and orchestration of several agents |
| Models | The documentation covers choosing a model for each agent and adding custom models |
| How it is priced | Published for the managed service: Essentials from $578 a month and Standard from $6,360 a month, described by IBM as indicative and varying by country. Premium is priced on request. No on-premises price is published |
| Limits to check | The documentation has pages on feature parity and on which prebuilt agents are available in each deployment, so expect differences between the managed service and the on-premises install and ask for the list |
Sources for IBM watsonx Orchestrate, read on 22 November 2026: product page [9], pricing [10], documentation [11].
Katonic AI
| Best for | Regulated organisations, governments and service providers that want agent building, a model gateway, knowledge, guardrails, evaluation and audit as one platform on infrastructure they own |
|---|---|
| Where it runs | On your premises on bare metal or an existing Kubernetes cluster, in your own cloud account, in a hybrid layout or air-gapped with offline installation bundles. It also runs on Red Hat OpenShift and SUSE Rancher |
| Who operates it | You or a local operator you choose. A service provider can run it for many customers under its own brand |
| Governance and audit | Four role tiers, approval by a person, and an audit trail across tenants, teams and tools. Every model call passes through one gateway where access rules, budgets and guardrails apply. Each agent version is scored against a test set before release. Katonic holds ISO/IEC 27001:2022 certification |
| How agents are built | Katonic Studio, with visual and code routes, and adapters for agents written in common frameworks |
| Models | Open-weight models served on your own GPUs, and any other model your IT team approves, behind the same gateway |
| How it is priced | Not published. Priced by quote |
| Limits to check | It is not a cloud, a model provider or a hardware vendor, so you supply the cluster, the GPUs and the models. It is a full platform with many services, which is more to install than a single-container tool. The price is not on the website |
Sources for Katonic AI, read on 22 November 2026: sovereign deployment [12], governance [13], trust centre [14], AI Gateway [15], AI Cloud [16].
Kore.ai
| Best for | Enterprises that want agents for customer service and employee work, built visually or in code in one studio |
|---|---|
| Where it runs | Software as a service by default. A self-hosted option is listed for enterprise customers, and the documentation names on-premises and private cloud deployment. Air-gapped operation and the install requirements are not stated on the pages we read |
| Who operates it | Kore.ai for the managed service. Not stated for the self-hosted edition |
| Governance and audit | The documentation lists single sign-on, access by role, customer-held encryption keys, guardrails, detection of personal data and audit logging, with approval steps in its workflow engine |
| How agents are built | A visual agent builder and a code route in the same studio, with several patterns for agents handing work to each other |
| Models | Commercial and open-source models, including models you have fine-tuned |
| How it is priced | Not published |
| Limits to check | Self-hosting is described for enterprise customers only. Ask what it installs on, whether it runs disconnected, and what it costs, since none of the three is on the public pages we read |
Sources for Kore.ai, read on 22 November 2026: agent platform [17], documentation [18].
Lyzr
| Best for | Organisations that want a private agent platform and may want it delivered as an appliance with GPUs included |
|---|---|
| Where it runs | A virtual private cloud on AWS, Azure or Google Cloud, your premises, or fully air-gapped. Lyzr also describes a portable on-premises appliance with GPUs that hosts open models, the agent builder and a fine-tuning tool |
| Who operates it | Hosted by Lyzr, in your cloud or on your premises, according to the pricing page |
| Governance and audit | Policies, guardrails and spend limits on each call, an owner and a release gate for each agent, approval by a person and audit trails. Lyzr states SOC 2 Type II, ISO 27001 and ISO 42001 certification |
| How agents are built | An agent studio for developers, a tool that turns a process described in plain language into an agent for IT to review, and prebuilt agents |
| Models | The model provider you already pay for, or open-weight models on your own hardware |
| How it is priced | Building and testing is free. Governing and automating are priced on request. A four-week proof of concept is listed at $20,000. The on-premises appliance is priced on request |
| Limits to check | No public price for on-premises or air-gapped use. Lyzr’s trust centre lists an outside model provider among its sub-processors, so ask which features need an outside service and which run fully inside the boundary |
Sources for Lyzr, read on 22 November 2026: home page [19], pricing [20], trust centre [21].
Onyx
| Best for | Teams whose first need is search and chat across company knowledge, with agents and actions on top, and who value an open-source core |
|---|---|
| Where it runs | Onyx Cloud, your own cloud account, bare metal or air-gapped. It installs with Docker or on your own Kubernetes cluster. The home page states it can run fully air-gapped with no outside third parties |
| Who operates it | You, for self-hosted installs. Onyx, for Onyx Cloud |
| Governance and audit | The Business plan lists access by role and inherited document permissions. The Enterprise plan adds SAML and OIDC single sign-on, group sync, custom roles and log export |
| How agents are built | Agents are set up with instructions, knowledge and actions. Actions connect through MCP or an OpenAPI file. There are developer APIs |
| Models | Any model, including open-weight models on your own GPUs |
| How it is priced | The Community Edition is free and open source. Business is $20 per user a month, billed yearly. Enterprise is priced on request, and the pricing page lists on-premises deployment under Enterprise |
| Limits to check | Code outside the enterprise folders is under the MIT licence, and the enterprise folders are under a separate Onyx licence. Guardrails and approval by a person are not named as features on the pages we read. Its strength is knowledge search, so check it against your needs for multi-step agents |
Sources for Onyx, read on 22 November 2026: home page [22], pricing [23], editions [24], licence [25].
VDF AI
| Best for | European organisations that want an on-premises agent platform from a vendor based in the European Union, with published cloud prices |
|---|---|
| Where it runs | A managed cloud, a private cloud, your premises, a sovereign cloud or air-gapped. It is delivered as signed container packages for any OCI-compatible runtime, and updates for disconnected sites arrive as signed files |
| Who operates it | VDF AI for its cloud. You for on-premises, with help from a VDF AI architect |
| Governance and audit | Access by role on every plan, a record of each prompt, tool call, retrieval and model route that can be sent to a security monitoring system, approval gates and a testing module. The trust centre maps the product to the EU AI Act, GDPR, DORA and NIS2 and says the mappings are not legal advice. No ISO 27001 or SOC 2 certificate is stated on the pages we read |
| How agents are built | Agents are created by chat or by form, with templates, a visual canvas for networks of agents, and APIs |
| Models | Your own choice of provider, local runtimes such as Ollama, and custom endpoints on your premises |
| How it is priced | Cloud prices are published: a free plan, Starter at $20 a month for up to three users, and Professional at $50 a seat each month. On-premises is sold as yearly seat bundles, with the list price on request |
| Limits to check | On-premises is licensed by seat, and API access is sold as separate expansion packs. Air-gapped deployment starts at the second bundle and includes a services engagement. |
Sources for VDF AI, read on 22 November 2026: home page [26], capabilities [27], pricing [28], trust centre [29].
Infrastructure platforms
Two products that sit under an agent platform. They serve models and run AI workloads on your hardware. Neither is sold as a place for business teams to build agents, and both are commonly paired with a product from the section above.
NVIDIA AI Enterprise
| Best for | Organisations that have bought NVIDIA GPUs and want supported software for serving models and the building blocks for agents |
|---|---|
| Where it runs | Certified data centre and edge systems, on bare metal, VMware vSphere, Red Hat OpenShift or the large public clouds. The documentation states that the NVIDIA Run:ai component supports connected and air-gapped deployment. Air-gapped support for the rest of the suite is not stated suite-wide on the pages we read |
| Who operates it | You, on your own hardware |
| Governance and audit | Policy and visibility for GPU use, security scanning of the containers and hardened containers for government use. Access control and audit for what staff do with a model are left to the layer above |
| How agents are built | NIM microservices for serving models, NeMo tools for evaluation and guardrails, and blueprints, which NVIDIA describes as example workflows. There is no builder for business users on the pages we read |
| Models | Open models packaged as NIM containers that run on your own infrastructure |
| How it is priced | Published in the licensing guide: a subscription at a list price of $4,500 per GPU for one year, or $1 per GPU an hour through cloud marketplaces, plus the cloud instance |
| Limits to check | It needs NVIDIA GPUs and is licensed per GPU. It is the base layer, not the place where a compliance officer or a claims handler builds and uses an agent |
Sources for NVIDIA AI Enterprise, read on 22 November 2026: product page [30], documentation [31], licensing guide [32].
Red Hat OpenShift AI
| Best for | Organisations that have standardised on Red Hat OpenShift and want model serving and AI operations on the same cluster |
|---|---|
| Where it runs | On your premises, at the edge or in the cloud, on Red Hat OpenShift. The product page states support for on-premises and disconnected mode |
| Who operates it | You. The product is self-managed |
| Governance and audit | A central hub for approved models, models offered as a service to developers, tracing for agents, vulnerability scanning and real-time guardrails on inputs and outputs |
| How agents are built | Operational tooling for agents: tracing, a catalogue of MCP servers and a playground for trying models. No agent builder for business users is described on the page we read |
| Models | Any model on any accelerator, according to Red Hat, including open-weight models and a set of validated third-party models |
| How it is priced | Not published on the product page. A 60-day trial is offered and needs an existing OpenShift cluster |
| Limits to check | It needs Red Hat OpenShift underneath. It is a platform for data science and platform teams, so plan for an agent layer on top. Red Hat’s documentation site refused automated reading on the day of our check, so this entry rests on the product page alone |
Sources for Red Hat OpenShift AI, read on 22 November 2026: product page [33].
- Which AI agent platform runs on Red Hat OpenShift? →
- Which agentic AI platform runs on HPE Private Cloud AI? →
- The NVIDIA Agent Toolkit on-premises and air-gapped →
Automation platforms with agent features
Two products that began as automation tools and now include agents. They suit work that is mostly a fixed process with a few steps that need judgement. For both, check which agent features exist in the edition you host yourself.
n8n
| Best for | Technical teams that want a visual workflow tool with agent steps, on their own servers, at low starting cost |
|---|---|
| Where it runs | n8n Cloud, or self-hosted on your premises or in a private cloud, with Docker and several other documented methods including OpenShift. The home page offers a fully on-premises option. Air-gapped operation is not stated on the pages we read |
| Who operates it | You, for self-hosted installs |
| Governance and audit | Approval by a person for tool calls is listed on every plan. Single sign-on and LDAP need the Business plan or above. Audit logging, log streaming and custom roles need the Enterprise plan |
| How agents are built | A visual builder, with code steps in JavaScript or Python where needed |
| Models | Any model, including open-source models |
| How it is priced | The self-hosted Community edition is free. Business, which is for self-hosted use, is €667 a month billed yearly with 40,000 workflow runs. Enterprise is priced on request |
| Limits to check | n8n is distributed under its own Sustainable Use License and an enterprise licence, not a standard open-source licence, so read the terms for your intended use. The pricing page says a paid self-hosted licence key contacts the n8n licence server each day, so ask about offline activation if the install must be disconnected. The free edition has no single sign-on |
Sources for n8n, read on 22 November 2026: home page [34], hosting documentation [35], pricing [36], repository [37].
UiPath
| Best for | Organisations with a large estate of UiPath robots that want agents inside the processes they have already automated |
|---|---|
| Where it runs | Automation Cloud, a dedicated cloud, a public sector cloud, or Automation Suite, which you host yourself on Kubernetes, including Red Hat OpenShift. The Automation Suite documentation lists offline installation for evaluation and for production |
| Who operates it | UiPath for Automation Cloud. You for Automation Suite |
| Governance and audit | Identity, access by role and tenant boundaries across the platform, with SAML, OIDC and SCIM. Tamper-evident logs of agent actions, prompts, tool calls and approvers, with export to a security monitoring system. Guardrails against prompt injection, and masking of personal data |
| How agents are built | Low-code in UiPath Studio, or code with the UiPath development kits |
| Models | Several commercial providers or customer-hosted models, through its AI Trust Layer. Bringing your own model is listed for the Enterprise plan |
| How it is priced | Basic starts at $25 a month and is cloud only. Standard and Enterprise, which include hosting on your own premises, are priced on request |
| Limits to check | UiPath calls Automation Cloud its recommended path. Its deployment options page compares features across the four deployment models and notes that in Automation Suite only the BPMN part of Maestro is available. Check that table for the agent features in the Automation Suite version you would install |
Sources for UiPath, read on 22 November 2026: deployment options [38], Automation Suite documentation [39], security and governance [40], pricing [41].
Developer frameworks
Three code libraries. Each is free to use and runs wherever your code runs, including on a disconnected network if the model is local. None is a platform. A framework gives an engineer the parts for one agent. It does not give the organisation access control, an audit trail, a place for staff to use the agent, or a way to stop it.
CrewAI
| Best for | Python teams that want to describe a group of agents with roles and have them work on a task together |
|---|---|
| Where it runs | The open-source framework runs wherever your Python code runs. The commercial product, CrewAI AMP, is offered on CrewAI’s cloud, in your own virtual private cloud or on your own infrastructure. Air-gapped operation is not stated on the pages we read |
| Who operates it | You, for the framework. CrewAI or you, for AMP |
| Governance and audit | The framework has no access control or audit of its own. AMP lists tracing, guardrails and approval by a person on both plans, and single sign-on, access by role and a dedicated private cloud on the Enterprise plan |
| How agents are built | Python code and configuration files, with a visual studio in AMP |
| Models | Many model providers, with OpenAI as the default. The repository documents local models through Ollama |
| How it is priced | The framework is free under the MIT licence. AMP has a free plan with 50 workflow runs a month, and Enterprise is priced on request |
| Limits to check | The framework sends usage telemetry unless you switch it off with a documented setting, which matters on a closed network. On-premises AMP is an Enterprise feature with no public price |
Sources for CrewAI, read on 22 November 2026: repository [42], pricing [43], AMP [44].
LangGraph
| Best for | Engineering teams that want precise control over an agent as a graph of steps with saved state |
|---|---|
| Where it runs | The library runs wherever your code runs. The commercial server around it, LangSmith Agent Server, is offered as a cloud service, as a hybrid with the control plane at LangChain and the agents in your infrastructure, self-hosted on your own Kubernetes cluster, or as a standalone server. The documentation names air-gapped use as a reason to self-host |
| Who operates it | You, for the library and for self-hosted servers. LangChain, for the cloud and for the hybrid control plane |
| Governance and audit | The library supports pausing for approval by a person. Tracing, evaluation, single sign-on and access by role come from LangSmith, with custom single sign-on and role controls on the Enterprise plan |
| How agents are built | Python or JavaScript code, with a visual studio for prototyping |
| Models | Any model provider |
| How it is priced | The library is free under the MIT licence. LangSmith is $0 a seat on the Developer plan and $39 a seat each month on Plus. Enterprise is priced on request |
| Limits to check | The library alone has no access control. Self-hosting with the control plane needs the Enterprise plan, and its price is not published |
Sources for LangGraph, read on 22 November 2026: product page [45], repository [46], deployment overview [47], pricing [48].
Microsoft Agent Framework
| Best for | Teams on .NET or Python that used AutoGen or Semantic Kernel and want the library Microsoft now develops in their place |
|---|---|
| Where it runs | Wherever your code runs. Microsoft describes it as the direct successor to AutoGen and Semantic Kernel, for .NET, Python and Go. The AutoGen repository was archived on 20 November 2026 and points new users to Agent Framework. Air-gapped use is not described on the pages we read |
| Who operates it | You |
| Governance and audit | Middleware, telemetry and tool approval are part of the library. A policy and audit feature set for agents is documented as experimental. Microsoft states that you are responsible for testing what you build and for your own safety measures |
| How agents are built | Code, with agents and graph-based workflows |
| Models | Microsoft Foundry, Azure OpenAI, OpenAI, Anthropic and others, and local models through Ollama |
| How it is priced | Free. No price is published because it is a library |
| Limits to check | It is a library, with no user interface for business teams and no control plane described. Most examples in the documentation use Azure services, so a team on a disconnected network should test the local model route early |
Sources for Microsoft Agent Framework, read on 22 November 2026: overview [49], model providers [50], security [51], AutoGen repository [52].
- AI agent frameworks compared: 9 options and how to choose →
- NeMo Agent Toolkit vs LangGraph, CrewAI and LangChain →
Looked at and left out
Eight more names were checked. Each was left out of the fifteen for a reason a reader can verify, not because of quality.
- Appian. Its pricing page says the platform can be deployed self-managed or on-premises, and a note on the same page says AI features may be unavailable or limited in self-managed deployments. Check agent support for your edition.
- Automation Anywhere. The AI Agent Studio page describes governance and guardrails. We could not confirm from the vendor’s pages that AI Agent Studio is available in an on-premises install.
- C3 AI. The documentation describes vendor-managed and customer-managed cloud deployment, and a single-node C3 AI Studio environment for on-premises and air-gapped clusters. Price is not published.
- deepset. The Haystack Enterprise Platform can be self-hosted on your infrastructure, and Haystack itself is an open-source framework. Air-gapped operation and price are not stated on the pages we read.
- Flowise. The open-source visual builder published a sunset notice with an end-of-life date of 31 August 2026. The code remains available under its licence.
- HPE Private Cloud AI. It is a turnkey private AI system and belongs in the infrastructure type. The product page would not open for our check on 22 November 2026, so it has no entry here. Two other guides on this site cover it.
- Microsoft AutoGen. Archived on 20 November 2026. Microsoft Agent Framework replaces it and is in the list.
- Prem AI. Its confidential inference service is hosted in Switzerland and can be licensed to run on-premises or in your own cloud. We could not find install details or prices for that option.
Sources for this section: Appian [53], Automation Anywhere [54], C3 AI [55], deepset [56], Flowise [57], AutoGen [52], Prem AI [58].
All fifteen side by side
The same facts in one table. "Stated" in the air-gapped column means the vendor’s own pages say the product, or the named edition, can run disconnected. "Not stated" means the pages we read do not say, not that it cannot.
| Platform | Type | Air-gapped | Where it runs and what it installs on | How it is priced | Source code |
|---|---|---|---|---|---|
| DataRobot | Agent platform | Not stated | Managed service, private cloud or on-premises, as a self-managed edition [2] | Not published [3] | Closed |
| Dify | Agent platform | Stated, Enterprise | Cloud, Docker for the free edition, Kubernetes for Enterprise [5] | Free edition. Cloud from $590 a workspace each year. Enterprise on request [7] | Open, with conditions |
| IBM watsonx Orchestrate | Agent platform | Confirm with IBM | IBM Cloud, AWS or on-premises on Kubernetes or IBM Software Hub [11] | Managed service from $578 a month. On-premises not published [10] | Closed |
| Katonic AI | Agent platform | Stated | On-premises, your cloud account, hybrid or air-gapped, on Kubernetes, OpenShift or SUSE Rancher [12] | Not published | Closed. Agent source code stays with the customer |
| Kore.ai | Agent platform | Not stated | Managed service by default. Self-hosted for enterprise customers [18] | Not published | Closed |
| Lyzr | Agent platform | Stated | Your cloud account, on-premises, air-gapped, or an appliance with GPUs [19] | Free to build. Proof of concept $20,000. The rest on request [20] | Closed |
| Onyx | Agent platform | Stated | Cloud, your cloud account, bare metal or air-gapped, on Docker or Kubernetes [22] | Free edition. Business $20 a user each month. Enterprise on request [23] | Open core |
| VDF AI | Agent platform | Stated | Managed cloud, private cloud, on-premises or air-gapped, as signed containers [28] | Cloud from free to $50 a seat each month. On-premises seat bundles on request [28] | Closed |
| NVIDIA AI Enterprise | Infrastructure | Stated for Run:ai | Certified systems on bare metal, vSphere, OpenShift or public cloud [31] | $4,500 per GPU a year at list price [32] | Closed, with open parts |
| Red Hat OpenShift AI | Infrastructure | Stated as disconnected | On-premises, edge or cloud, on Red Hat OpenShift [33] | Not published | Open-source based |
| n8n | Automation | Not stated | Cloud or self-hosted with Docker and other methods [35] | Free edition. Business €667 a month. Enterprise on request [36] | Source available |
| UiPath | Automation | Stated as offline install | Cloud, or Automation Suite on Kubernetes that you host [39] | From $25 a month in the cloud. Self-hosted on request [41] | Closed |
| CrewAI | Framework | Not stated | Library anywhere. AMP in the cloud, your private cloud or your infrastructure [43] | Library free. AMP Enterprise on request [43] | Open, MIT |
| LangGraph | Framework | Stated for self-hosted server | Library anywhere. Agent Server as cloud, hybrid, self-hosted or standalone [47] | Library free. LangSmith from $0 to $39 a seat each month. Enterprise on request [48] | Open, MIT |
| Microsoft Agent Framework | Framework | Not stated | Library anywhere your .NET, Python or Go code runs [49] | Free | Open |
How to choose
Work through these in order. Each step removes names from the list.
- Decide the type. If you have no agent layer, you need an agent platform. If you have GPUs and no way to serve models, start with infrastructure. If the work is a fixed process with a few judgement steps, look at automation. If you have engineers and one agent to write, a framework may be enough for now.
- Write down the boundary. On-premises with an outside connection, or fully disconnected. If it is disconnected, keep only the products whose vendor states it for the edition you would buy, and ask the others in writing.
- Check the edition, not the brand. Several products in this list keep single sign-on, audit logs or on-premises deployment for the top plan. Compare the edition you can afford, not the feature list of the most expensive one.
- Read the licence. Open source can come with conditions. A licence may restrict running the software as a service for other organisations, or require a daily licence check. Both matter for a service provider and for a closed network.
- Ask who operates it. If the vendor’s engineers run the system inside your account, they are part of your risk assessment. If your staff run it, they need the skills and the hours.
- Ask for the on-premises price. Most of the fifteen do not publish one. Ask what the price scales with: users, runs, GPUs or tenants. A price per user suits a small expert team. A price per GPU or per platform suits a service used by the whole organisation.
- Test with your own work. Take one real task, your own documents and an open-weight model you are allowed to host. Run the same test on two or three products. Score the answers before anyone sees a demonstration.
- Plan the exit. Ask how you take your agents, your data and your logs with you. An agent you hold as source code can move. An agent that exists only as settings in a vendor’s builder may not.
Running models on your own GPUs also has an energy cost that a cloud invoice hides. Lead Green, a Katonic project, publishes a public index of GPU energy per million output tokens, which helps when you size hardware for the models you plan to host [60].
By industry
The same list reads differently in each sector, because the rule that forces the choice is different.
- Bank. A bank usually needs an agent platform on its own infrastructure, with a record of every model call and a person approving each decision. Shortlist the agent platforms whose vendors state on-premises or air-gapped operation, and ask each how its audit trail is exported to your own monitoring.
- Insurer. An insurer handles policy wordings, claims files and health data. Look for document permissions that carry through to the agent and for approval steps before anything is sent to a customer.
- Hospital. A hospital needs patient data to stay inside its network and a clinician to decide. Open-weight models on local GPUs, access by role and a full audit trail matter more than the number of connectors.
- Telecom operator. An operator often wants to run the platform for itself and also sell it to enterprise customers. That needs separate tenants and a licence that allows it, so read the multi-tenant terms of each product first.
- Government agency. An agency with classified or restricted work needs a product that runs with no outside connection, with updates carried in as signed files. Keep only the vendors that state this in writing.
- Manufacturer. A manufacturer often runs agents close to plants with weak links to head office. Check that the platform runs at a site with its own models, and that it reaches the systems on the plant network.
- Energy utility. A utility separates its corporate network from its operational network. Agents that read from operational systems belong on infrastructure the utility controls, and no agent should change a control system.
- Law firm. A law firm must keep each client’s documents apart from every other client’s. Look for matter-level permissions, a record of what each agent read, and a deployment the firm hosts itself.
- Logistics company. A logistics company runs on many partner systems and a lot of documents. A platform with strong connectors and approval steps for customs and billing work suits it better than a framework.
Where Katonic fits
Katonic AI is in the agent platform type. It is the layer for building, running and governing AI agents on infrastructure you own, and it sits above an infrastructure platform such as Red Hat OpenShift AI, NVIDIA AI Enterprise or a turnkey private AI system. It runs on Kubernetes or Red Hat OpenShift, works with open-weight models you host, and can run with no outside connection.
It is a reasonable first choice when you want the gateway, knowledge, agent building, guardrails, evaluation and audit from one vendor, when the system must run disconnected, or when a provider wants to offer the platform to its own customers under its own brand. It is not the right choice if you only need one agent written in code, where a framework costs less, or if you want a vendor to host everything for you in its own cloud.
Read next
- Top sovereign AI platforms in 2026: a buyer evaluation guide →
- Compare AI agent platforms: every comparison on this site →
- AI agent frameworks compared →
- What is a private AI cloud? →
- AI agent use cases that run on your own infrastructure →
Questions and answers
Which AI agent platform can a bank run on its own infrastructure?
Several. Agent platforms whose vendors state an on-premises edition include DataRobot, Dify Enterprise, IBM watsonx Orchestrate, Katonic AI, Kore.ai, Lyzr, Onyx and VDF AI. A bank should keep the ones that give it an audit trail of every model call, approval by a person and, if required, operation with no outside connection, and then test two or three with its own documents.
Which AI agent platforms can run fully air-gapped?
On 22 November 2026 the vendors of Dify Enterprise, Katonic AI, Lyzr, Onyx and VDF AI stated air-gapped operation on their own pages. LangChain names air-gapped use for its self-hosted Agent Server, UiPath documents offline installation of Automation Suite, and Red Hat states a disconnected mode for OpenShift AI. For the others, ask the vendor in writing.
What is the difference between on-premises, self-hosted and air-gapped?
On-premises says where the system runs: in a facility you control. Self-hosted says who operates it: you. Air-gapped means it runs with no connection to any outside network. A system can be on-premises and self-hosted and still depend on an outside service for models, updates or a licence check.
Can a hospital or an insurer use a hosted AI agent service instead?
Sometimes. It depends on the rules for the data involved and on what the regulator and the organisation’s own risk team accept. Where patient or policyholder data may not leave the organisation’s network, the agent platform and the models have to run inside it.
Is an open-source framework such as LangGraph or CrewAI enough?
For one agent written and run by an engineering team, it can be. A framework does not provide access by role, an audit trail, a place for staff to use the agent or a way to govern many agents. Organisations usually add a platform for those once more than one team is building.
Do I need NVIDIA AI Enterprise or Red Hat OpenShift AI as well as an agent platform?
You need something to serve models on your GPUs. NVIDIA AI Enterprise and Red Hat OpenShift AI are two supported ways to get it. An agent platform sits above that layer and is where agents are built, governed and used. Some agent platforms bring their own model serving, so check before buying both.
How much does an on-premises AI agent platform cost?
Most vendors do not publish an on-premises price. Of the fifteen here, published figures cover cloud plans, free editions, one infrastructure licence at $4,500 per GPU a year and one self-hosted plan at €667 a month. Ask each vendor what the price scales with and budget separately for hardware, models and the people who will run it.
Which platform lets a telecom operator or a government agency serve many organisations?
One that separates tenants and whose licence allows it. Katonic AI describes a multi-tenant edition that a provider runs under its own brand. Dify’s open-source licence does not allow a multi-tenant service on its source code without written permission. For any other product, ask about tenant separation and resale terms before a pilot.
Sources
Facts checked on 22 November 2026 against each vendor’s own pages. Statements about Katonic AI come from this website.
- DataRobot, home page: https://www.datarobot.com/
- DataRobot, AI Platform: https://www.datarobot.com/product/ai-platform/
- DataRobot, trial page: https://www.datarobot.com/trial/
- DataRobot documentation, Self-Managed AI Platform: https://docs.datarobot.com/11.1/en/docs/index.html
- Dify Enterprise: https://dify.ai/dify-enterprise
- Dify documentation, self-hosting overview: https://docs.dify.ai/en/self-host/deploy/overview
- Dify pricing: https://dify.ai/pricing
- Dify licence: https://github.com/langgenius/dify/blob/main/LICENSE
- IBM watsonx Orchestrate: https://www.ibm.com/products/watsonx-orchestrate
- IBM watsonx Orchestrate pricing: https://www.ibm.com/products/watsonx-orchestrate/pricing
- IBM watsonx Orchestrate documentation: https://www.ibm.com/docs/watsonx/watson-orchestrate/base
- Katonic AI, sovereign deployment: https://www.katonic.ai/sovereign
- Katonic AI, governance: https://www.katonic.ai/governance
- Katonic AI, trust centre: https://www.katonic.ai/trust
- Katonic AI Gateway: https://www.katonic.ai/ai-gateway
- Katonic AI Cloud: https://www.katonic.ai/ai-cloud
- Kore.ai, agent platform: https://www.kore.ai/ai-agent-platform
- Kore.ai documentation: https://docs.kore.ai/
- Lyzr, home page: https://www.lyzr.ai/
- Lyzr pricing: https://www.lyzr.ai/pricing/
- Lyzr trust centre: https://security.lyzr.ai/
- Onyx, home page: https://www.onyx.app/
- Onyx pricing: https://www.onyx.app/pricing
- Onyx documentation, editions: https://docs.onyx.app/admins/getting_started/enterprise_edition
- Onyx licence: https://github.com/onyx-dot-app/onyx/blob/main/LICENSE
- VDF AI, home page: https://vdf.ai/
- VDF AI, on-premises platform capabilities: https://vdf.ai/on-prem-ai-platform-capabilities/
- VDF AI pricing: https://vdf.ai/pricing/
- VDF AI trust centre: https://vdf.ai/trust-center/
- NVIDIA AI Enterprise: https://www.nvidia.com/en-us/data-center/products/ai-enterprise/
- NVIDIA AI Enterprise documentation: https://docs.nvidia.com/ai-enterprise/
- NVIDIA AI Enterprise licensing guide, pricing: https://docs.nvidia.com/ai-enterprise/planning-resource/licensing-guide/latest/pricing.html
- Red Hat OpenShift AI: https://www.redhat.com/en/products/ai/openshift-ai
- n8n, home page: https://n8n.io/
- n8n documentation, hosting: https://docs.n8n.io/hosting/
- n8n pricing: https://n8n.io/pricing/
- n8n repository: https://github.com/n8n-io/n8n
- UiPath, deployment options: https://www.uipath.com/platform/agentic-automation/deployment-options
- UiPath Automation Suite documentation, overview: https://docs.uipath.com/automation-suite/automation-suite/2.2510/installation-guide/automation-suite-overview
- UiPath, security and governance: https://www.uipath.com/platform/agentic-automation/foundation/security-governance
- UiPath pricing: https://www.uipath.com/pricing
- CrewAI repository: https://github.com/crewAIInc/crewAI
- CrewAI pricing: https://www.crewai.com/pricing
- CrewAI AMP: https://crewai.com/amp
- LangGraph: https://www.langchain.com/langgraph
- LangGraph repository: https://github.com/langchain-ai/langgraph
- LangSmith documentation, deployment overview: https://docs.langchain.com/langsmith/deploy-overview
- LangChain pricing: https://www.langchain.com/pricing
- Microsoft Agent Framework overview: https://learn.microsoft.com/en-us/agent-framework/overview/
- Microsoft Agent Framework, model providers: https://learn.microsoft.com/en-us/agent-framework/integrations/by-component/model-providers/
- Microsoft Agent Framework, security: https://learn.microsoft.com/en-us/agent-framework/agents/security
- Microsoft AutoGen repository: https://github.com/microsoft/autogen
- Appian pricing: https://appian.com/products/pricing
- Automation Anywhere, AI Agent Studio: https://www.automationanywhere.com/products/ai-agent-studio
- C3 AI documentation, about the platform: https://docs.c3.ai/docs/platform/8.12/topic/about-platform
- deepset, Haystack Enterprise Platform: https://www.deepset.ai/haystack-platform
- Flowise, sunset notice: https://flowiseai.com/sunset
- Prem AI, Enclave: https://www.premai.io/enclave-api
- The Sovereignty Test, canonical version: https://premnaraindas.com/sovereignty-test/
- Lead Green, a public index of GPU energy per million output tokens: https://leadgreen.ai/