Eclipse Launches Enclave 1.0 to Isolate AI Agents

The Eclipse Foundation launches Eclipse Enclave 1.0, a sandboxed environment for running programming agents with defined permissions. Compatible notably with Claude Code and Codex, the platform regulates their access without promising to eliminate risks.
Eclipse Does Not Promise Total Security
Sandboxing does not exempt one from controlling the software produced. Thabang Mashologu, CMO and product lead at the Eclipse Foundation, emphasizes that Enclave establishes technical boundaries outside of AI but does not eliminate risk. He presents it as a means of managing risk, rather than a guarantee of complete safety.
The official considers the development based on autonomous agents to be still in its early stages. The permissions granted, network connections, and dependencies remain to be examined by developers. The quality and security of the generated code must also be verified.
Access is Governed by Explicit Permissions
In Enclave, agent sessions run in isolated containers, with restricted access to sensitive resources. According to Thabang Mashologu, developers can select project sources, documents, web resources, and tools necessary for the task; SSH keys, tokens, and unauthorized files remain inaccessible. Built-in protections also apply to API keys and OAuth tokens used with third-party services.
Sessions can operate without administrative privileges using Docker or Podman. Their traffic passes through a gateway that enforces an allowlist, while a log keeps track of the destinations the agents attempted to reach. In addition to Claude Code and Codex, Enclave supports OpenCode, GitHub Copilot CLI, Pi, and Eclipse Theia; its architecture can accommodate other tools.
Git repositories provide direct access to tools and features. A separate extensions repository offers Antigravity, additional isolation functions, and Playwright MCP. A VNC viewer is used to manipulate graphical interfaces inside the container; extensions can also add commands like enclave vnc-viewer.
The distribution includes binaries for Linux and macOS, including for arm64, as well as deb and rpm packages. On Windows, a launcher relies on WSL2. Automatic updates have stable and continuous channels, with versions available on the Eclipse download site.
Le brief IA que les pros lisent chaque soir
Les 7 actus IA du jour, décryptées en 5 min. Gratuit.
Inclus dès l'inscription : notre sélection des meilleurs guides & comparatifs IA.
Choisissez votre rythme
Gratuit · Pas de spam · Désabonnement en 1 clic
Developers Retain Validation of Changes
Programming agents are no longer limited to suggestions: they can execute commands, modify software, run tests, or install dependencies. With unlimited rights or minimal oversight, these capabilities expose systems to unauthorized changes, credential leaks, and unintended actions. Conversely, requiring approval for each operation significantly reduces productivity gains, according to Thabang Mashologu, who refuses to pit productivity against security.
Refactoring illustrates this compromise, he explains: an agent may need to traverse the code, modify it, install dependencies, test, analyze logs with scripts, and start over. These extended tasks can save time but complicate unrestricted access to the machine. At EclipseSource, where daily adoption has been rapid, repeated validations previously interrupted developers and stalled agents when no one responded immediately, the official reports.
According to Thabang Mashologu, defining access upstream according to company policies has reduced this oversight and facilitated parallel delegation. Developers could better prepare objectives, architecture, constraints, context, tools, and expected verifications. However, they still review changes before merging, focusing more on outcomes than on each intermediate operation.
Enclave allows multiple simultaneous agents, whose isolated work facilitates review. Mashologu cites a bug fix, writing tests, and refactoring conducted separately, each in its own work tree. The model remains at the developer's discretion, and persistent sessions retain progress and configuration during a resume. The platform also includes a preview version of HomeShell, an interface designed to organize parallel sessions of a project.
The Foundation Advocates for Independent Infrastructure
For the Eclipse Foundation, the issue goes beyond technical isolation: it concerns control over the infrastructure of agents. Thabang Mashologu advocates for systems that would not be captive to any vendor, model, or development tool. As autonomy increases, he believes that the mechanisms governing access, authorized commands, and external interactions become essential, complementing the power of the models.
The foundation, a non-profit organization recognized in the European Union, supports and governs over 450 open-source projects. Enclave was initially designed in late 2025 for the internal needs of EclipseSource, a member that develops custom tools and integrated development environments. Present in open source for over 15 years, the company brought the platform to the foundation in July 2026.
Its team continues development and gradually transfers functions from its internal tools. According to Mashologu, the openness of the code gives others the opportunity to contribute, inspect the technology, and participate in evolution choices.
Brief IA — L'actualité IA en français
L'essentiel de l'actualité de l'intelligence artificielle, décrypté et expliqué chaque jour.