IBM Made Its AI Coding Agent Air-Gappable. It's the Fourth Company This Week to Make the Same Bet.

IBM made its agentic coding platform, Bob, generally available as a fully self-hosted, on-premises, or air-gapped deployment on September 30. Taken alone, that’s a solid enterprise feature launch. Taken next to this week’s other coverage on this site — AWS testing EU operational independence, Cloudflare building isolated per-agent storage, Nvidia shipping hardware-level agent containment — it’s the fourth distinct company this week making the identical underlying bet: the current competitive edge in enterprise AI isn’t more capability, it’s more control over where that capability actually runs.

What IBM actually shipped

Bob retains its core experience across deployment modes — the BobShell IDE, parallel tool calling, skills, and operating modes all carry over, so developers aren’t trading functionality for containment. Three deployment configurations are supported at general availability: fully self-hosted (the customer installs and manages the model on their own infrastructure), air-gapped, and hybrid, where Bob connects to external model services instead.

At GA, only two models are supported for the fully self-hosted configuration: NVIDIA Nemotron and Poolside Laguna. That’s a real, specific constraint worth knowing before you evaluate this — if your organization is committed to a different model family, the self-hosted path doesn’t yet support it, and you’d need the hybrid configuration instead, which reintroduces the external-connection dependency self-hosting is meant to avoid.

Optional premium packages extend Bob to Java modernization work and IBM i and IBM Z platform support — a notable detail for enterprises running legacy mainframe or midrange systems that aren’t well served by most AI coding tools built around modern web stacks.

The caveat nobody’s headline mentions: no published pricing

IBM has not published pricing for the self-hosted deployment. Analysts covering the announcement were direct about why that matters: the self-hosted option’s actual success depends on customer willingness to pay for deployment flexibility against real integration and support costs, not just on whether the capability exists. IBM’s stock moved up nearly 6% on the announcement — a real market signal that investors read this as evidence of IBM’s traction in security-focused enterprise AI — but a stock bump on a feature announcement and sustained paid adoption of an unpriced product are two different things worth tracking separately over the coming months.

IBM isn’t alone — this is now a four-way field

A direct comparison is worth laying out, because self-hosted or air-gapped agentic coding has gone from rare to nearly standard among major vendors within what looks like a single year:

• IBM Bob (self-hosted) — air-gapped with supported self-hosted models (Nemotron, Laguna)
• GitLab Duo Agent Platform Self-Hosted — air-gapped available with an offline license
• Mistral Vibe for Code (Enterprise) — uses open-weight Devstral models, which inherently suit air-gapped use since nothing depends on a hosted API
• GitHub Copilot CLI (BYOK) — offline mode available with a local model

Four different vendors, four different technical approaches, the same destination. If your organization has been waiting for “the” air-gapped coding agent to emerge as a clear category winner, that’s not what’s happening — what’s happening is the capability becoming table stakes across the field, which means the actual differentiator is shifting to model quality, premium integrations (like IBM’s Java/IBM Z packages), and total cost, not whether self-hosting exists at all.

The pattern across this whole week, stated plainly

Reading this site’s last several days together: AWS is testing whether its European cloud can run without its own global backbone. Cloudflare is building storage specifically so AI agents never need to touch a real developer’s git configuration. Nvidia shipped a hardware-isolated layer specifically so an agent’s own oversight system can’t be the thing the agent persuades or tampers with. IBM just made its flagship coding agent capable of running somewhere that never talks to IBM’s servers at all.

None of these four companies coordinated this messaging — the timing is coincidental in the narrow sense that none of these launches reference each other. But the underlying driver is the same, and it’s not coincidental at all: a full month of disclosed agent incidents (Anthropic, Gemini, OpenAI, the GitSpawn and Plugin4Shell vulnerabilities) has made “can this agent’s activity be contained to infrastructure we actually control” the question enterprise buyers are now asking before they ask what the agent can do. Containment has become the pitch, not an add-on to it.

What to actually do if you’re evaluating this

• Confirm your required model is on the self-hosted-supported list before assuming IBM Bob’s air-gapped mode fits your stack — Nemotron and Laguna only, at GA.
• Price out the real integration and support cost, not just the deployment flexibility, given IBM hasn’t published self-hosted pricing — ask directly rather than assuming parity with the hosted tier.
• Compare against GitLab, Mistral, and GitHub’s equivalent offerings on your actual workload, not just on whether “self-hosted” is checked — the real differences are in model support, legacy-platform coverage, and total cost, which vary meaningfully across all four.

Official reference: IBM’s own newsroom announcement (ibm.com/new/announcements) and SiliconANGLE’s coverage, cross-checked against MarkTechPost’s vendor comparison and AIwire’s reporting, all dated September 30 – October 2, 2026. Self-hosted model support and pricing are the kind of detail likely to expand post-GA — confirm current status directly with IBM before planning a deployment around specifics here.