Modernization
Self-Hosted IBM Bob for IBM i: Where Does Your Code Actually Go?
A practical look at self-hosted IBM Bob for IBM i: code context, model endpoints, OpenShift requirements, IBM i authority, and what to validate before a proof of concept.
Self-hosted IBM Bob gives an organization control over the Bob backend. To understand where its IBM i source code goes, the team also needs to identify the model endpoint, the context sent to it, and the tools connected to the development session.
For an IBM i team, interest in an AI development tool can stall on a very practical question: can we use it with our application source?
That source might contain decades of pricing rules, customer-specific processing, interface contracts, and operational knowledge. A useful RPG explanation does not settle whether the organization can send that information to a particular service.
IBM’s October 1, 2026 announcement gives teams another deployment option. It dates general availability of self-hosted Bob to September 24 and confirms that the IBM i premium package works with that deployment. The Bob backend runs on customer-managed Red Hat OpenShift. IBM release announcement
The important design decision is where inference runs: the processing that turns your prompt and context into a model response.
This article is an architecture explainer based on IBM documentation reviewed on October 7, 2026. It does not report a completed self-hosted installation or performance benchmark.
Follow the context from the developer to the model
Consider a developer asking Bob to explain an order-entry program before changing it. The architecture has four places worth identifying:
| Part | What to establish for your deployment |
|---|---|
| Developer environment | Which workspace, source members, and tool results can the session access? |
| IBM i connection | Which partition and user profile are used to read objects or perform actions? |
| Bob backend | Which OpenShift environment operates the Bob services? |
| Model endpoint | Which system processes the prompt and any code context included in the request? |
IBM explicitly says that, with the cloud-model route, requests including code context go to the organization’s cloud account. Keeping the backend inside your network therefore does not, by itself, keep model processing there. IBM release announcement
The following is a conceptual view of those relationships, not a packet-level description of every Bob service:
IBM i application
^
| Code for IBM i / SSH
v
Developer's Bob IDE
|
| Prompt and selected context
v
Customer-managed Bob backend on OpenShift
|
| Model request
+----> Approved cloud model endpoint
|
+----> Customer-operated model endpoint
(alternative inference route)
The review should trace information in the returned context as well as the original source. For example, a developer may ask for a program explanation and then ask a connected tool for database information. Whether those results contain only metadata or include business data depends on the task and tool use. That is a reason to inspect representative sessions during the proof of concept.
Choose the model route deliberately
IBM documents model endpoints hosted in public cloud services, on separate private inference infrastructure, or on the OpenShift cluster using OpenShift AI. A working model endpoint is a prerequisite; Bob IDE does not provision or operate model-serving infrastructure for you. Model serving documentation
| Route | What the team needs to resolve |
|---|---|
| Customer-managed backend with a cloud model service | Whether the selected endpoint, region, contractual terms, and handling of code context meet the organization’s requirements. |
| Customer-managed backend with private inference | Who operates the model, how capacity is sized, and whether the selected model performs adequately on the team’s RPG and SQL work. |
| Disconnected deployment | How the approved environment receives installation artifacts and updates, and how the complete workflow behaves without external services. |
For a team already using an approved model service, the first route may offer a practical starting point. For a team whose code must remain entirely within a defined private environment, the private-inference route deserves evaluation.
Those are architectural judgments, not claims that either arrangement automatically meets a particular policy. The answer depends on the actual endpoints, configuration, and operating practices.
IBM i support still needs a separate backend platform
IBM’s current requirements specify amd64/x86-64 OpenShift nodes for the Bob backend. Supported OpenShift versions listed at review time are 4.20, 4.21, and 4.22. An IBM i partition is the application target in this design; its presence does not satisfy the backend requirement. System requirements
Sizing also has several layers. IBM separates the resource requests of Bob workloads from the surrounding OpenShift platform and explicitly excludes model-serving resources from its Bob sizing estimates. A quoted Bob Core footprint should therefore not be presented as the total infrastructure needed for private AI inference. System requirements
A dedicated cluster is not mandatory: IBM documents coexistence with other workloads when sufficient resources are available. The organization takes responsibility for operating the backend, including lifecycle management and the configuration of networking, storage, and identity. Self-hosted overview
This changes the conversation for a small IBM i team. If a platform team already operates suitable OpenShift and model infrastructure, the work may fit an existing service. If those capabilities do not exist, evaluating Bob also means evaluating a new operational responsibility.
The connected IBM i profile matters
IBM’s security guidance says the i premium package communicates through Code for IBM i over SSH, using the active connection’s user profile. Operations available to that profile are available to Bob through the connection. IBM i security guidelines
Moving the backend does not change the authority you have assigned to that IBM i profile. For a first evaluation, I would use a dedicated development environment and a profile with the access required for the chosen task. If the task is explaining a program, the evaluation should not quietly expand into unrestricted production administration.
The same review should cover the temporary libraries and IFS locations used by the development tools. IBM also calls out access to Understand’s stored scan data: sharing it can expose information about objects that another user cannot access directly. IBM i security guidelines
These are separate questions from model hosting. A private model can still be paired with excessive application access.
Include operating work in the evaluation
IBM’s installation guide covers cluster prerequisites, container registries, model endpoints, configuration, certificate trust, and user onboarding. It assigns deployment and maintenance work to the cluster administrator. Installation overview
I would ask the platform team to demonstrate a normal developer sign-in and a failed connection during the evaluation. Both matter. A tool is easier to adopt when the team knows who handles an expired certificate, an unavailable endpoint, or an access problem.
There are also release-specific limitations worth accounting for. At review time, IBM documents that Activity Logs are unavailable in the Admin UI, controlled in-place upgrades are unsupported, and Bobalytics is unavailable in the self-hosted stack. A fresh installation is recommended for a new release; hotfixes have a separate supported path. Known limitations
The overview assigns security-event logging and monitoring to the OpenShift platform. The infrastructure team therefore needs an explicit plan for collecting, retaining, and reviewing the required logs. Self-hosted overview
Make the proof of concept answer a real question
In the earlier Era of i Bob benchmark, useful explanations and plausible architecture could coexist with implementation mistakes. A different hosting arrangement creates a new environment to evaluate; it does not establish the quality of the selected model on your applications.
I would start with one synthetic IBM i application and a small set of repeatable tasks: explain a program, identify a known SQL defect, and suggest a limited change. Preserve the prompts, model identity, versions, tool actions, and unedited first responses.
Before expanding the evaluation, ask the team to produce five concrete answers:
- Information flow: Which endpoints receive source, context, and tool results during these tasks?
- Application access: Which IBM i profile and libraries are involved, and which actions can it perform?
- Engineering quality: Are the explanations correct, and do the proposed changes compile and behave as intended?
- Operational ownership: Who supports the backend, model endpoint, certificates, logs, and upgrades?
- Cost and capacity: What are the measured response times and the expected platform, model, and support costs for the intended team size?
Self-hosted Bob is worth considering when deployment boundaries are preventing an otherwise useful development workflow. The strongest case for adopting it will combine an acceptable information flow with demonstrated IBM i usefulness and a team prepared to operate the service.
References
Documentation and supporting sources for this article.

Comments
Share your thoughts, questions, or real-world IBM i experiences related to this article.