Product2 publishers3 min readPublished
IBM's self-hosted Bob coding agent runs air-gapped on Nvidia and Poolside models
IBM's Bob coding agent now runs air-gapped inside customer networks on two supported local models, Nvidia's Nemotron and Poolside's Laguna. Banks can keep their code in-house, though fully local work runs on a shorter model list than the Claude, Mistral and Granite mix IBM pitches for Bob.
The Product Desk · Product desk
What happened
- IBM launched Bob as a cloud service in April and said at the time that on-premises deployment would follow in a later release.
- In a supported self-hosted setup, source code, development context and build artifacts stay inside the customer-managed environment, IBM said.
- A hybrid option keeps Bob on the customer's own infrastructure while it connects to supported external model services.
- The self-hosted version is generally available and keeps the BobShell IDE, parallel tool calling, skills and operating modes, according to IBM.
- IBM's mainframe add-on, the Bob Premium Package for Z, gets the same self-hosted option plus deeper Z-specific code analysis.
Compiled by The Product DeskSomething wrong?How this is made
Why it matters
- constraint Teams that want a model beyond Nemotron or Laguna have to use the hybrid path, so part of their work leaves the network and needs its own compliance sign-off.
- cost The customer runs the model, so local inference hardware and model licenses come out of the bank's own budget unless existing licenses qualify under IBM's bring-your-own terms.
- decision Before rollout, each repository needs a residency label, because a single Bob install can send one codebase to a local model and another to an external service.
IBM's own example customer is a developer at a bank. That developer works on core banking software using a model that runs on the bank's own hardware. For less restricted work, they connect to an approved external service, and Bob works the same way in both setups [17]. The person running the developer platform at that bank now has to keep two configurations straight and show compliance which repository went where.
Devops.com describes Bob as routing tasks across Anthropic's Claude, Mistral's open models and IBM's Granite, choosing by accuracy, performance and cost [3]. A self-hosted install needs a model the customer supplies, and for models on the customer's own hardware IBM currently supports Nvidia's Nemotron and Poolside's Laguna [14]. Granite, IBM's own model family, is missing from that list as SiliconANGLE reported it [14].
According to devops.com, reluctance to send source code to someone else's cloud has slowed adoption at banks, insurers and government agencies [18]. "Organizations need AI that operates inside environments they have control over, especially when working with sensitive code and regulated data," said Neel Sundaresan, IBM's general manager of AI and automation [7]. For evidence of demand, IBM points to its own survey. In a June Institute for Business Value report, 68% of executives said meeting data residency and sovereignty requirements across geographies is challenging [9]. That question covers every kind of data in every region. Neither report measures whether residency outweighs agent quality as the reason regulated teams hold back.
Vendors tend to assume their internal numbers will carry over to customers. More than 80,000 IBM employees use Bob, up from 100 developers in June 2025 [4], an 800-fold rise [1]. Surveyed users report an average productivity gain of 45% [5]. A seat count at the company selling the tool measures distribution. A self-reported survey measures how people rate their own speed. A bank's developers will be running Nemotron or Laguna on code that has never left the building. The useful measures there are time to the first merged change and how much of the agent's output gets rewritten.
Mainframes are where IBM's bet matters most. Devops.com wrote that much of the code enterprises most want to modernize runs on mainframes, and that it is the code they are least willing to expose [19]. "Modernizing mainframe code behind the firewall at banks, insurers, and government agencies is the work that gets IBM in the door," said Mitch Ashley of The Futurum Group [10]. He added: "An agent trusted within those core systems is positioned to become the platform on which the same teams build their own agents. That makes choosing Bob a multi-year commitment." [11]
I'd run a pilot on the local model before signing. The cost is a slower start, and the self-hosted option meets the compliance requirement whether the pilot goes well or badly. For each codebase, two tests decide it. First, does its residency rule permit any call to an approved external service? Second, does Nemotron or Laguna do acceptable work on it in the pilot? Code that has to stay inside and passes the pilot is the case IBM built air-gapped Bob for [1]. If the code has to stay inside and the pilot fails, the team ends up compliant with an agent nobody uses. Code that is allowed out can go hybrid [6]. For that code, the self-hosted option matters less than which external model the bank has approved.
What to watch
- Whether IBM adds models beyond Nemotron and Laguna, Granite in particular, to the list it supports on customer hardware.
- Whether customers use Bob for modernization, new agent development, or both, the test Futurum's Mitch Ashley set.
- Customer-reported results from an air-gapped deployment, set beside IBM's internal 45% survey figure.