Incident sheet · /security

Filed
Read
4 min · 830 words
Severity
Routine

Eclipse launches Enclave 1.0 sandbox for autonomous coding agents

Eclipse Enclave 1.0 gives six coding tools isolated containers, scoped files and credentials, network allowlists, parallel worktrees and persistent sessions while developers retain code review.

AI coding agents need a safe place to work; Eclipse is building the sandbox to give it to them
FeedAI coding agents need a safe place to work; Eclipse is building the sandbox to give it to them — AI-generated

Event log

  1. Eclipse Enclave 1.0 supports six coding tools: Claude Code, Codex, OpenCode, GitHub Copilot CLI, Pi and Eclipse Theia.
  2. The Eclipse Foundation supports and governs more than 450 open-source projects.
  3. Enclave runs rootless sessions on Podman and Docker, with all traffic passing through an allowlist gateway.
  4. EclipseSource began Enclave in late 2025 and contributed it to the Foundation in July 2026.
  5. The platform ships Linux and macOS binaries, including arm64, plus deb and rpm packages and a Windows launcher through WSL2.

By Amara Osei

Eclipse Enclave 1.0 launches as a vendor-neutral sandbox for six supported coding agents, with isolated containers and explicit controls over files, credentials and network access.

What does Enclave change?

AI coding tools now move beyond suggestion-only copilots toward autonomous agents that execute terminal commands, modify software, install dependencies and run tests.

Agents with little oversight can make unauthorized changes, expose credentials or sensitive data, and take unintended actions, as vendors have discovered.

The nonprofit Eclipse Foundation, chartered in the EU, supports and governs more than 450 open-source projects. It launched Enclave 1.0 as a vendor-neutral sandbox and runtime for autonomous coding tools.

The alternative keeps agents on a “very short leash,” with developers approving every action and losing much of autonomy’s productivity benefit.

Thabang Mashologu, Eclipse Foundation CMO and head of products, said: “Ultimately, developers shouldn’t have to choose between agent productivity and security.”

Takeaway: Enclave narrows agent access while preserving independent work.

Which agents and models can use it?

Version 1.0 supports Claude Code, Codex, OpenCode, GitHub Copilot CLI, Pi and Eclipse Theia. Its extensible architecture allows more tool integrations.

Developers run sessions in isolated containers. Built-in mechanisms protect API keys and OAuth tokens for third-party services while limiting access to sensitive resources.

Users grant only required project source files, selected documentation, web resources and curated development tools. SSH keys, unrelated tokens and other source files stay out of reach.

In a substantial refactoring, an agent can explore the codebase, make changes, install dependencies, run tests, execute ad hoc scripts, analyze runtime logs and iterate until completion.

Mashologu said: “That’s exactly the kind of work where agents can save developers a lot of time,” while noting that long-running tasks make unrestricted, unsupervised access uncomfortable.

Takeaway: Control follows the task rather than exposing the whole machine.

How does Enclave control access and traffic?

  • Rootless session operation runs on Podman and Docker.
  • Network traffic from every session passes through an allowlist gateway, and logs record what the agent tried to reach.
  • Git integration exposes tools and features directly from repositories. A separate enclave-extensions repository adds harnesses such as Antigravity, sandbox features and developer tooling such as Playwright MCP. A VNC viewer operates UIs inside the container, and extensions can add commands such as enclave vnc-viewer.
  • Binaries cover Linux and macOS, including arm64, with deb and rpm packages; Windows users launch it through WSL2.
  • Self-updates use stable and rolling channels, and Eclipse publishes releases on its download site.

Takeaway: Enclave makes access and outbound traffic inspectable instead of leaving agents with implicit access.

Can agents work in parallel?

Multiple agents can run parallel tasks in separate environments, making their work easier to review. Developers can choose models and use persistent sessions that preserve progress and configuration.

A preview of HomeShell, a user interface for managing multiple parallel sessions, organizes parallel work in project workstreams.

One agent can fix a bug, another write tests and a third refactor code. Each uses its own worktree and isolated environment without interfering with the others, Mashologu explained.

Developers still review code before merging. They can define changes, assess results and prepare context and checks rather than supervise every step.

Takeaway: Parallel execution shortens queues without merging security boundaries.

Where did Enclave come from?

EclipseSource, a Foundation member that builds custom tools and integrated developer environments, started Enclave in late 2025 as an internal tool.

Developers there adopted it daily, turning it from a theoretical exercise into working infrastructure.

Repeated approval prompts interrupted their flow, and agents stalled whenever they could not respond immediately. Defining access rules upfront in line with company policies reduced that babysitting.

Developers then focused on outcomes, architecture, constraints, tools, context and verification checks. That structure also made parallel delegation more practical.

EclipseSource, active in open source for more than 15 years, contributed Enclave to the Foundation in July 2026. Its team continues developing the platform and porting features from internal tooling.

Open-sourcing lets others contribute, scrutinize the technology and shape its direction.

Takeaway: A working internal tool became governed open-source infrastructure.

What risk remains?

Agentic development remains at an early stage. Sandboxing defines boundaries outside the AI harness and contains risk, but it does not remove risk.

Developers still must assess permissions, network access, dependencies, and the quality and security of code agents produce.

Mashologu said Enclave does not make autonomous agents completely safe; it gives developers better tools to manage risk.

The larger policy issue is control of the infrastructure agents rely on. That layer should remain open and independent of any single vendor, AI model or development tool, Mashologu said.

As agents gain autonomy, infrastructure must govern what they can access, execute and expose to the outside world. Model power alone cannot replace trusted infrastructure.

Takeaway: The sandbox reduces exposure; developers still carry final responsibility.

Enclave moves governance from each agent command to the access policy around the session.

via infoworld.com/feed/ (Source)

Filed under

  • ai-coding-agents
  • sandboxing
  • eclipse-foundation
  • container-isolation
  • autonomous-agents
Share this article:

More from Amara Osei

Amara Osei

Show full bio

Correspondent covering regulation, policy and their impact on enterprise information technology and the leaders who run at CIO Monitor.

203 articles

The newsroom digest

One email a day with the stories that matter. No noise.

By subscribing you agree to the privacy policy.

Cross-references

‹ Previous articleNext article ›