IBM Bob is IBM’s AI coding agent for enterprises: an IDE and a terminal tool, called Bob-Shell, that take a task and route it to whichever model fits, Claude for hard work, Mistral’s open models or IBM’s own Granite for the rest, so nobody on the team picks a model. It went generally available on 28 April 2026, and since October 2026 it also runs inside your own data center or on an air-gapped network. That last part is why it exists: Bob is for the banks, insurers and public bodies whose code may not leave the building, and whose auditors want one log of what the agent did.
What IBM Bob is
| IBM Bob | |
|---|---|
| Maker | IBM |
| Interfaces | Bob IDE; Bob-Shell in the terminal |
| Models | Claude, Mistral and IBM Granite, chosen per task by Bob’s router; local and private model options on-premises |
| Where it runs | IBM-managed SaaS, or on-premises including customer-controlled and air-gapped environments |
| License | Proprietary. Granite, the model, is Apache-2.0; Bob, the agent, is not open source |
| How you pay | Per instance, in tiers from a free trial to Enterprise; usage is metered in “Bobcoins”; paid add-ons for Java, IBM i, IBM Z and DevSecOps |
| Scope | Planning, design, coding, testing, deployment and modernization, with governance and security controls, per IBM |
The add-ons are the tell for who Bob is for. Java, IBM i and Z are the languages and systems of long-lived enterprise software; the Z package, a private preview, adds an Architect mode that maps dependencies and change impact before anything is edited.
What’s different from Claude Code or Copilot
- The model is chosen for you. Every other agent on this site makes you pick a model, or defaults to its vendor’s. Bob routes: lighter models for simple edits, stronger ones for multi-step work. IBM says this cuts AI compute spend by about 40%, an IBM figure.
- It runs where the code already is. Claude Code reaches your data center through Bedrock, Vertex or Foundry; Copilot’s agents run on GitHub’s machines. Bob’s on-premises build runs the agent and the models behind your firewall, with no internet needed.
- Governance is the product, not a setting. Policy, audit and security scanning are built in, which is what a regulated company buys it for. IBM screens for risky packages and runs deeper scans on demand.
- It’s priced per instance, with add-ons. Copilot and the Claude and ChatGPT plans are priced per user. Bob’s model is closer to enterprise software, and the premium packages are extra.
What to know before you buy
- It was breakable before GA. In January 2026, during the preview, researchers showed the terminal tool could be manipulated into running malware and the IDE was open to the usual data-exfiltration tricks against AI tools. IBM says both are addressed; it’s the right question to put to the sales team.
- The 45% productivity gain is IBM’s own number, measured on IBM’s own staff across multi-step workflows. Treat it as a vendor claim until your pilot measures something.
- No benchmark row. Bob has no entry on Terminal-Bench or SWE-bench, so it can’t be compared on a leaderboard with Claude Code and Codex. Its router uses the same frontier models, which is the stronger argument anyway.
- The name is a joke IBM made on purpose. Yes, Microsoft Bob. IBM knows.
IBM Bob alternatives
| Tool | What it is | Inside the firewall? | Pick it if |
|---|---|---|---|
| GitHub Copilot Enterprise | Autocomplete, chat, terminal agent and cloud agent, with organization policies | Partly: the code lives on GitHub; the agents run on GitHub’s machines | Your code is already on GitHub and you want one bill |
| Claude Code | Anthropic’s agent, through Bedrock, Vertex or Foundry for enterprise accounts | Through your cloud account, not air-gapped | You want the top benchmark rows and already run a cloud |
| Codex | OpenAI’s agent, open source, with a native Windows sandbox | Through Azure or an OpenAI enterprise agreement | You standardized on OpenAI |
| OpenCode with local models | Open-source terminal agent with Ollama, vLLM or llama.cpp | Fully, with your own GPUs | You want air-gapped and open source, and can run models |
| Contenox | Open-source desktop app with a policy engine that decides what the agent may run | Fully, with local models through Ollama or vLLM | You want rules you can read, per project, on Windows |
The license, model and sandbox matrix for all of these is in the AI coding harness comparison.
Contenox
Contenox covers the part of Bob that isn’t the model router: what runs without asking, what asks, what’s blocked, as rules you can read in Permissions and keep in your repository. It runs with a local model through Ollama or vLLM when the code can’t leave the building, and with a ChatGPT plan or an API key when it can. Download it here.
Related guides
- Best AI coding agents: where Bob sits, and why
- AI coding harness comparison: 16 agents, every cell checked
- AI agent permissions: what to allow, ask and block
- Run local LLMs for coding: the air-gapped route without IBM