aixplain Launches Koder to Bring Agentic Coding Directly to Devices

AI coding agents are becoming good at writing, modifying and reviewing code - very good. But as they become more a part of the software development process, they are also raising a less straightforward question: where does all that work happen?

For developers and organisations working with sensitive code, proprietary data or strict security and data-sovereignty requirements, the answer to that question is critical. Cloud-based AI tools can create challenges when source code and development context are anchored in controlled environment.

It’s a question San Jose-based AI infrastructure specialist aixplain is addressing with Koder - an intelligent coding agent that runs locally on a developer’s device. Founded in 2020 by Saudi-German entrepreneur, angel investor and tech expert Dr. Hassan Sawaf, aixplain launched Koder at LEAP x DeepFest 2026 in Riyadh as part of its broader focus on building, deploying and governing agentic AI, with Arabic-first capabilities and sovereign AI deployment at its core - and it's a potential game-changer.

What Makes Koder Different?

Koder takes a local-first approach, allowing source code, context, work history and inference processes to remain on the developer’s device unless they explicitly choose to use a cloud-based model.

That's particularly relevant as coding agents move beyond generating individual snippets and start handling more complex, multi-step development tasks. In layman's terms, the more an agent can do, the more of a developer’s environment it may need to access. For organisations handling sensitive information, the cloud is very often not an option.

“No developer should have to choose between the smartest AI and the safety of their own code, and with Koder no one has to again,” said Dr. Hassan Sawaf, Founder and CEO of aixplain. “It puts a world-class coding agent on your own machine: your code stays yours, your secrets stay secret, and you choose the model, the environment, and the price. From a student on a laptop, to a government inside an air-gapped network, the best AI tools now come to you, on your terms. That is what freedom and sovereignty in AI look like.”

Local execution also means Koder can operate without a conventional cloud connection. It can be used offline, inside secure facilities and on completely air-gapped networks, without network latency or usage-rate limits. Local inference also eliminates token-based billing, which can become significant as agentic coding workloads consume substantially more resources than conventional conversational AI.

In addition, developers are not tied into a particular model or infrastructure. They can use a local or custom model, deploy Koder on-premises or in an isolated environment, or access an advanced cloud model without changing their workflow or compromising their security requirements. When cloud-based models are used, Koder provides access to the aixplain marketplace through a single API key and unified billing, giving teams a consolidated view of costs, usage and performance.

What Can Koder Actually Do?

Koder is designed to support the software development lifecycle rather than simply generate code. Working from the command line, developers can use the agent for planning and implementation, code changes and review, pull request creation, and the deployment, evaluation and refinement of aixplain agents. It can work across multiple repositories in a single session, maintain context as developers move between tasks and allow them to revert to previous states without losing that context.

Koder can also use multiple agents to tackle complex problems in parallel and solve them faster. The system also supports MCP servers and thousands of integrations, allowing Koder to connect with tools already embedded in developers’ workflows. It can therefore sit within an existing development stack rather than requiring developers to rebuild their workflow around the agent.

How Does Koder Keep Developers in Control?

Giving an AI agent more responsibility also raises questions about oversight. If an agent can plan, coordinate and execute increasingly complex tasks, developers need visibility into what it's doing and the ability to understand or audit its actions. Koder is built on aixplain’s agentic operating system and incorporates governance, observability and access controls, including step-by-step telemetry and an auditable activity log within the workspace.

“One agent today can write or change thousands of lines of code before lunch, and the real challenge isn't speed; it's giving agents that power without losing control,” said Dr. Nur Hamdan, Chief Product Officer at aixplain. “We designed Koder to be the control layer for agentic development: agents plan, coordinate, and execute while every decision remains policy-bound, traceable, and auditable, and developers keep control of quality, policy, and delivery. Koder is for a new generation of developers who will design and direct intelligent systems, applying human judgment where it matters most.”

The distinction is important as coding agents take on more responsibility: greater automation does not have to mean less visibility for the developer.

How Does Koder Handle Different Models?

Koder allows developers to switch models during a session without restarting, giving them the flexibility to match different stages of a project with different capabilities.

Its 'Auto Mode' can select a model for each step based on cost and performance. According to aixplain, this approach aims to reduce cloud spending by around 30% while maintaining quality. Rather than committing to a single model throughout a project, developers can use Auto Mode to match individual tasks with models according to their cost and performance requirements.

What Could This Mean for the Future of Coding?

As coding agents move from tools that suggest code to systems capable of planning tasks, coordinating multiple steps and making substantial changes across software projects, the conversation around them is changing too.

Capability is only one part of the equation.

Where the work happens, what data an agent can access and how much control developers retain are becoming increasingly important questions - particularly for organisations working with sensitive source code or strict data-sovereignty requirements.

For MENA, those questions are particularly relevant as governments and enterprises continue to explore AI infrastructure that meets local security, regulatory and data-sovereignty requirements. In that context, the ability to run an agent within a controlled environment becomes less a technical preference than a consideration in whether certain organisations can adopt agentic coding at all. Koder is currently available through the command line, with aixplain stating that installation takes less than 60 seconds. Local use is free, while cloud-based models are billed through aixplain credits.

But Koder’s potential extends beyond developers. As AI lowers the barrier to building, its applications can reach designers, musicians and other creatives looking to turn an idea into reality.

With AI coding agents taking on more of the development process, the question is no longer simply what they can build. It is also where they can work, what they can access and who gets to build with them.