Eclipse Enclave 1.0 Puts Your AI Coding Agent in a Sandbox The Eclipse Foundation released Eclipse Enclave 1.0 on 19 November 2026, a free, MIT-licensed, vendor-neutral sandbox that runs coding agents in isolated containers with only the files, network hosts and secrets a user explicitly allows. Built from internal tooling at EclipseSource and contributed to the Foundation in July 2026, Enclave ships six built-in agents, uses a gateway sidecar to filter DNS and proxy outbound connections with a readable audit log, and passes credentials into containers only as placeholders. It runs on Linux and macOS with Docker or Podman backends and a WSL2 launcher for Windows. Coding agents stopped being autocomplete a while ago. They run shell commands, install packages, edit files and run tests by themselves now, and the useful ones do it for long stretches without asking you anything. That autonomy is the whole point. It is also the problem: on a normal workstation, the same agent that refactors your code can also read your SSH keys, leak a token, or wipe the wrong folder. The Eclipse Foundation's answer, announced 19 November 2026, is Eclipse Enclave 1.0. A free, open-source, MIT-licensed sandbox that gives each agent its own isolated container, with only the files, network hosts and secrets you choose. I haven't run Enclave hands-on yet — this is a launch-day piece, built from the Foundation's announcement, the project docs, and independent coverage — so take the setup details as per the docs, not as my war story. But the design decisions here are worth talking about, because they are the decisions every agent sandbox has to get right, and Enclave gets most of them right on the first try. Every coding agent vendor now ships some kind of sandbox. The problem is that each one is tied to its own agent, its own cloud, its own rules. Switch agents and you start over. Enclave is the vendor-neutral answer: one runtime, six agents built in, and an architecture that lets more be added. It came out of real internal use at EclipseSource, which built the tooling for its own teams and contributed it to the Eclipse Foundation in July 2026. That origin story matters — this wasn't designed in a slide deck. The pitch is simple: let the agent run wild inside the box, and keep the box small. Enclave splits the difference between "watch the agent's every command" and "hand it the whole machine." Inside the container, the agent runs with no confirmation prompts — the project calls it YOLO mode. Outside it, the host stays out of reach except for paths you mount on purpose, usually the project worktree. The four headline capabilities, per the Foundation's announcement: Six tools are built in: Claude Code, Codex, OpenCode, GitHub Copilot CLI, Pi and Eclipse Theia, which can attach to a running session. Amp and Antigravity ship separately in an enclave-extensions repository, and tooling can be pulled directly from git repos. The agent landscape changes every quarter. A sandbox tied to one vendor's agent would be obsolete by spring. A container alone is a weak sandbox if the agent inside it can phone home anywhere. Enclave's answer is a gateway sidecar that filters DNS lookups and proxies outbound connections. Only domains on your allowlist get through, and both allowed and blocked attempts are written to a log you can read with enclave network log . That log is the audit trail: what the agent tried to reach, not just what it reached. Secrets get special handling too. Credentials the gateway can add to requests reach the container only as placeholders, so the real token never sits where the agent can print it or upload it. Enclave also refuses to start under the root account, and it blocks mounts of your home folder and similar sensitive locations. Small decisions, but they are the decisions that stop the classic "agent pasted the API key into a log file" accident. For teams, the audit side may matter more than the isolation. Gateway logs give a record of what each agent tried to reach, and the project says it plans to grow this toward evidence for the EU Cyber Resilience Act and the EU AI Act. A 451 Research report cited in the announcement makes the same enterprise case. Linux and macOS, including arm64, with deb and rpm packages, plus a Windows launcher via WSL2. The container backends are Docker and Podman — Podman works rootless, Docker needs rootful mode with buildx. Binaries self-update on stable or rolling channels with enclave self-update . You need Git 2.26 or later with an author and committer identity set; without it, Enclave stops before starting the container. A graphical companion called HomeShell is available as a preview — it groups sessions into projects and workstreams and lets you watch and answer waiting agents from a web UI or desktop window, though it has no stable release yet. Now the caveats, because a launch-day piece that hides them is just a press release with better formatting. First, the dates. The public launch announcement came on 19 November 2026. The Eclipse project registry dates version 1.0.0 to 31 October 2026. So the code existed before the announcement, and the announcement is the launch moment, not the first release date of the code itself. Say "announced 19 November" and you are accurate. Say "released 19 November" as if the code appeared that day and you are not. Second, the QEMU backend. There is an experimental QEMU microVM backend, and it currently operates without the Docker gateway's network restrictions. If you run Enclave on that backend, do not assume the allowlist is protecting you. It is not, on that backend, yet. The container backends are where the gateway actually enforces. And the caveat that applies to every sandbox, not just this one: a sandbox limits damage, it does not judge code quality. Anything the agent writes still needs review, and an agent with database credentials can still run bad SQL against whatever server you allow it to reach. Point it at a development schema only. If you already let an agent run on your laptop with broad permissions, Enclave is a low-effort way to shrink the blast radius. Mount only the repository, keep the network allowlist tight, and review the agent's branch before you merge it. The worktree model is handy for parallel work too: hand one agent a refactor and another the test scripts, then compare both branches without them stepping on each other's files. Download it from enclave.eclipse.dev https://enclave.eclipse.dev . After install, change to a project folder and run enclave as a normal user. On the first launch you pick an agent, and Enclave remembers that pick in its config file. From then on, enclave --tool picks another agent for one run, enclave continue resumes the last session, and enclave ps lists what is running. Enclave joins a wider push to isolate agents, alongside efforts such as Google's gVisor and the Agent Substrate donation to the CNCF. The difference here is scope: not a cloud service, not a single vendor's product. A local runtime you install, under a foundation, with an MIT license. Nobody can take the sandbox away from you later. Are you already letting a coding agent run unsupervised on your machine — and if you tried a sandbox, which one, and what made you trust it? Sources: Eclipse Foundation announcement via GlobeNewswire, 19 November 2026; Eclipse Enclave project documentation and README; independent coverage of the launch, 20 November 2026. The Eclipse project registry dates 1.0.0 to 31 October 2026 while the public announcement came 19 November 2026; the experimental QEMU microVM backend operates without the Docker gateway's network restrictions. No hands-on testing was performed for this article. Originally published on ToolsDig https://toolsdig.com/posts/eclipse-enclave-1-0-ai-coding-agent-sandbox.html .