Library

Bob, watsonx Orchestrate and the IBM i MCP Server: How the Chain Actually Fits Together

Updated September 2, 2026

Three products, three different jobs, and exactly one link in the chain that IBM has never documented. That is the short version, and it is worth being precise about which link, because the August 2026 watsonx Orchestrate release produced a run of headlines that made it sound like IBM had finally wired its AI platform into IBM i.

It did not. It wired two build tools together. Which is genuinely useful, and is also not the same thing.

The three pieces and what each one is for

IBM Bob builds

IBM's AI coding agent for IBM i. Bob 2.0 and the Bob Premium Package for i went generally available June 24, 2026. Bob writes and modifies code, and since August 2026 it also writes watsonx Orchestrate agents, tools and workflows. It is a development tool. It does not run in production.

watsonx Orchestrate runs

IBM's agent and workflow platform. It hosts the agent, calls its tools, handles the conversation, and as of August 2026 monitors the result through AgentOps, generally available on IBM Cloud and AWS. It is general-purpose. It has no idea what a physical file is.

The IBM i MCP Server opens the door

An open-source, IBM-published Model Context Protocol server, in beta since October 2025. It runs on IBM i, reaches Db2 for i through Mapepire, and exposes SQL tools defined in YAML plus IBM i services to any MCP-aware agent. This is the only piece that actually touches your data.

Say those three jobs out loud in order and the design assembles itself. Bob builds the agent. Orchestrate runs the agent. The MCP Server is what the agent calls when it needs to know what is in a file on your system. No custom middleware, no vendor in the middle.

The reason this needs an article at all is that the three pieces were built by different teams for different reasons, and only two of the three were ever meant to meet.

What is real today

The Bob to Orchestrate link is real and supported. In the Create an agent dialog, Build with Bob launches Bob into a workspace already connected to your Orchestrate tenant. You describe what you want in plain language and Bob produces the agent definition, the tools it calls and the workflow, and it will refine an agent that already exists rather than only producing first drafts. Under the covers it configures two MCP servers of its own, one that searches the Orchestrate Agent Development Kit documentation and one that drives the Orchestrate command line.

That second detail is why this chain is worth taking seriously rather than filing under speculation. Bob already consumes MCP servers as a matter of routine. Writing a tool definition that points at one more MCP server is not a stretch for it. It is the thing it does.

One correction to the headlines, and it matters for planning. IBM's own release notes list Build with Bob as a preview in the Create an agent dialog, on IBM Cloud and AWS deployments. The Orchestrate extension on the Bob side is generally available. The Launch Bob entry point inside Orchestrate is not. AgentOps reaching GA on IBM Cloud and the removal of the one-knowledge-source-per-agent limit, both in the same release, are GA. If someone hands you a slide saying Build with Bob shipped GA in August, the slide is wrong.

The MCP Server side is real too, with its own asterisk. It is open source, published by IBM on GitHub, and still labeled beta. It does not train anything and it is not a machine learning feature. It is a connectivity layer that turns named SQL statements and IBM i services into tools an agent can call. IT Jungle read IBM's May 2026 direction the same way: Db2 for i has no vector store, OpenSearch on Power is IBM's search and vector layer, and the MCP Server is the agent's door into IBM i. Consistent story, told across three separate announcements.

What is still yours to build

The middle link. An Orchestrate agent calling the IBM i MCP Server is a design, not a documented integration. Every piece exists and is published, but nobody at IBM has written the walkthrough that puts all three together against a Db2 for i workload, and as of September 2026 no IBM material connects Build with Bob to IBM i at all.

That is not a reason to avoid it. It is a reason to size the work honestly. Here is what nobody hands you:

Authority modelWhat user profile the IBM i MCP Server runs under, and what that profile can reach. Decide this before an agent ever calls it, not after.
Tool scopeNamed, parameterized SQL tools in YAML beat open-ended SQL access. Every tool you expose is a thing an agent can be talked into calling.
Network routeHow a SaaS Orchestrate tenant on IBM Cloud or AWS reaches an MCP server running inside your data center. This is usually the piece that eats the schedule.
Failure behaviorWhat the agent does when the query times out, the subsystem is down, or the answer is empty. Agents are confident. Your job is to make sure they are not confidently wrong.
Support storyThree products, two of them beta or preview, one integration nobody documented. Know who you call before you need to call them.

Who this is for, and who should wait

This chain is worth piloting if you already have a real process that spans IBM i and something else. Order entry on IBM i with a cloud CRM on the other side. Dispatch on IBM i with a customer portal somewhere newer. The value of an agent is that it crosses the boundary between the two, and without that boundary an agent is an expensive way to run a query you could have written in SQL.

It is also worth piloting if you have a platform team with cloud experience and a real reason to be in watsonx Orchestrate anyway. In that case Build with Bob is a productivity feature on top of a decision you already made.

Wait if you are a two-developer shop. Not because the technology is beyond you, but because the return is not there yet. The IBM i AI investment that pays at that size this year is Bob with the Premium Package for i pointed at the programs nobody understands any more. IBM's July 15, 2026 Power announcement cites Heartland Co-Op estimating 60 percent faster time for developers new to the platform to understand complex applications with the Premium Package. That is a customer estimate inside an IBM release, not an independent benchmark, and it is still the most concrete productivity number published for AI on this platform. Nothing comparable exists for the agent chain, because almost nobody has run it yet.

Wait too if your reason for looking at this is that a vendor showed you a slide with all three logos on it. The slide is not wrong. It is just further ahead of the documentation than it looks.

The order to do it in

A sane build order for the Bob, Orchestrate and IBM i MCP Server chain

StepWhat you are provingStop if
1. Bob with the Premium Package for i, on real codeThat AI-assisted work on your codebase returns something useful, before any of this touches an agent platformYour developers do not trust the output. Fix that first; the rest gets harder, not easier.
2. IBM i MCP Server, one read-only YAML SQL toolThat an MCP client can ask a plain-language question and get a correct answer from Db2 for iYou cannot answer the authority question cleanly. Do not proceed on a shrug.
3. A watsonx Orchestrate agent calling that toolThat the middle link works at all, in your network, with your tenantThe network route to your data center turns into a project of its own. That is fine, just budget it separately.
4. Build with Bob to generate the next agentThat the productivity story holds once the design is provenNothing. This step is the easy one, and it is a preview, so keep a manual fallback.
5. Anything that writes to IBM iThat you have a reason strong enough to justify itYou have not run the read-only version in production for a meaningful stretch first.

Sources: watsonx Orchestrate release notes, August 2026; IBM announcement on Bob and watsonx Orchestrate, August 19, 2026; IT Jungle, May 4 and July 13, 2026; IBM Newsroom, July 15, 2026.

IBM TechXchange runs October 26 to 29, 2026 in Atlanta, the same event where Project Bob was first shown in 2025. If IBM is going to document this chain properly, or ship a prebuilt IBM i skill for watsonx Orchestrate, that is the likeliest venue. Until it does, this is an assembly job with three moving parts, two of which carry a beta or preview label. Worth building. Worth building carefully.

Sources

Related Tools & Vendors

More From the Library