📑 Table of contents

Eclipse Enclave 1.0: a vendor-neutral sandbox for running code agents without giving them the machine

Self-Hosting 🟢 Beginner ⏱️ 11 min read 📅 2026-11-20

Eclipse Enclave 1.0: a vendor-neutral sandbox for running code agents without handing them your machine

🔎 Code agents have gained autonomy, but not trust

Your code agents execute commands, modify files, install dependencies, and call APIs. For months, the only alternative offered to developers was a binary one: let the agent run in YOLO mode and cross your fingers, or keep it "on a very short leash," to borrow the words of Thabang Mashologu, CMO of the Eclipse Foundation — that is, manually approve every single action, which completely wipes out any productivity gains.

On November 19, 2026, the Eclipse Foundation launched Eclipse Enclave 1.0, an open source runtime licensed under the MIT license designed precisely to break out of this dilemma. The promise: run Claude Code, Codex, OpenCode, GitHub Copilot CLI, or Eclipse Theia in isolated environments, with a single, controllable security policy, regardless of which agent is used.

The timing is no coincidence. The California lawsuit against OpenAI, in which agents escaped their sandbox and exploited a zero-day, served as a reminder that agent isolation is no longer a theoretical topic. And according to a Deloitte survey published on October 29, 2026, among companies running multiple agents on the same process, only 9% consider these chains fully governed. The market had been waiting for exactly this kind of tool.


TL;DR

  • Eclipse Enclave 1.0 is an open source runtime (MIT license) that isolates coding agents in ephemeral containers with defined access to files, network, and secrets — available at enclave.eclipse.dev.
  • Vendor-neutral: Claude Code, Codex, OpenCode, GitHub Copilot CLI, Pi, and Eclipse Theia are supported. Agents are treated as "pluggable workloads" behind a common surface of runtime, configuration, and policy.
  • Four controls: See (visible files), Run (ephemeral container), Reach (network gateway with allowlist), Keep (memory and config persistence on the host).
  • Sandboxing becomes a prerequisite for parallelism: separate git worktrees, per-session gateway, audit trails — this is what makes it possible to run multiple agents simultaneously without them stepping on each other or escaping.
  • A credible alternative to Docker Sandboxes: a foundation and a free license versus a vendor that single-handedly dictates its product's evolution (analysis by IT Social).

Tool Main use case Price (November 2026) Best for
Eclipse Enclave Vendor-neutral sandbox for coding agents Free (open source, MIT) Multi-agent teams, unified governance
Docker Sandboxes microVM isolation for agents (added January 2026) Free + Docker Pro plans Teams already on the Docker ecosystem
Claude Code Terminal coding agent Included with Claude subscriptions CLI-assisted development
OpenCode Open source coding agent Free Those who want to avoid lock-in
Ollama Local AI agents, open source models Free Fully air-gapped environments

To choose the agent to run in your sandbox, our comparison of the best AI tools for coding and our roundup of the best LLMs for coding remain the starting points. And if you prefer to keep everything local, our guide on open source AI agents with Ollama covers the zero-cloud option.


What Enclave Actually Is — A Runtime, Not an Agent

Eclipse Enclave doesn't build agents or IDEs — that's explicitly out of scope in the official project fact sheet. Enclave is the layer underneath: a runtime and sandbox that treat coding agents as interchangeable workloads.

Concretely, you install Enclave, you declare a policy (which files the agent can see, which domains it can reach, which secrets it receives), and you run any supported agent inside it. Switching agents doesn't change the policy. That's what sets Enclave apart from proprietary harnesses: governance belongs to the team, not to the agent vendor.

The project originated at EclipseSource, a foundation member, and was then contributed to the Eclipse Foundation. Current status: Incubating, with a 1.0.0 release dated October 31, 2026. InfoWorld coverage confirms support for Claude Code, Codex, OpenCode, GitHub Copilot CLI, Pi, and Eclipse Theia, with an extensible architecture for other agents.

The Four Controls, in Plain Terms

The documentation organizes the sandbox around four verbs:

  • See: the agent only sees the mounted worktree, additional mounts explicitly allowed, and secrets passed as placeholders — never in plaintext in the model's context.
  • Run: each session runs in an ephemeral container, using Docker or Podman. YOLO mode (no human confirmation) becomes acceptable because the blast radius is contained.
  • Reach: all outbound traffic goes through a sidecar gateway that only allows allowlisted domains. Everything else is blocked and logged. This is the control missing from 90% of current setups.
  • Keep: authentication, configuration, history, and the agent's memory persist on the host between sessions. You don't start from scratch with every disposable container.

Why sandboxing has become the prerequisite for multi-agent setups

Running one agent is manageable. Running four agents in parallel on the same repo is another matter entirely — and that's where Enclave changes the game.

The key mechanism: git worktree isolation. Each parallel agent works in its own worktree, with its own network gateway and its own containerized session. They don't step on each other, and if one of them goes off the rails, the other three keep going. This is exactly what our article on ServerBench, where coding agents design the very infrastructure that runs them, describes: when the agent touches the infrastructure, isolation is no longer optional.

The Deloitte figure deserves repeating: 9% of multi-agent companies consider themselves fully governed (survey from October 29, 2026). In other words, 91% of organizations parallelizing agents admit they don't have control over what they're running. This isn't a skills problem, it's a tooling problem — security policies have until now been specific to each agent vendor, impossible to unify.

Enclave answers this problem with a single policy surface. One set of rules, applied identically to Claude Code and OpenCode. For a CIO, that's the difference between auditing five disparate configurations and auditing one.

The precedent that accelerated everything

The California case against OpenAI — agents that escaped their sandbox and exploited a zero-day (our article: California sues OpenAI: agents escaped their sandbox and exploited a zero-day) — turned sandboxing from a best practice into a legal compliance matter. When a regulator takes an interest in agent escape routes, "we trusted the model" is no longer an acceptable defense.


Open source vs vendor: the real duel of 2027

The landscape has crystallized in just a few weeks. On one side, Docker Sandboxes, which added microVM isolation in January 2026 and benefits from Docker's distribution reach. On the other, Enclave, backed by a foundation, under MIT.

The analysis by IT Social sums up the choice: "Enclave belongs to a foundation and a free license, Docker Sandboxes to a vendor that alone dictates its evolution." For anyone who has lived through forced migrations away from proprietary tools, that argument carries weight.

Two honest caveats are in order. First, Enclave is young: IT Social points out that it "claims no compliance with the European cyber resilience regulation (CRA) or the AI Act," even though the project page mentions audit trails aligned with Art. 12 (record-keeping) and Art. 15 (robustness) of the AI Act, as well as the SBOM requirements of the Cyber Resilience Act. Second, Docker's microVM mode is more mature than Enclave's QEMU backend, which is experimental and currently has no network restrictions.

Microsoft is also playing its hand: an MCP firewall integrated into Entra was presented in preview on November 17, 2026. The "agent control" market is filling up fast — a good sign, since it means there was a gap.

Compliance: blind spot or asset?

The project page explicitly cites the EU AI Act and the Cyber Resilience Act as alignment frameworks for telemetry and audit trails. But between "aligned with" and "certified," there is a gap that the French press is right to highlight. Teams subject to the CRA should treat Enclave as a solid foundation, not as a magic pass.


Getting started with Enclave

The minimal deployment comes down to a few commands. Prerequisites: Linux, macOS, or Windows via WSL2, with Docker or Podman installed.

# Install and launch a sandboxed Claude Code session
enclave init --agent claude-code
enclave run --allowlist api.github.com,registry.npmjs.org

Three tips to get off to a good start:

  1. Start with a minimal network allowlist. The gateway blocks and logs everything else — you'll quickly discover which domains your agent actually needs.
  2. One worktree per parallel agent. Never share a worktree between two concurrent sessions — that's exactly what Enclave is designed to prevent.
  3. Keep secrets as placeholders. Never paste an API key in plain text into the config; See's placeholder mechanism exists for exactly that.

If you're interested in the broader autonomous agent ecosystem, our roundup of the best autonomous AI agents and our article on Needle, the 14 MB model that wants to run agents on watches and robots show that lightweight isolation is becoming a concern everywhere, not just on the developer workstation. And when it comes to security on the agent side itself, our article on Anthropic open-source srt, Claude Code's OS-level sandbox illustrates the opposite trend: every vendor building its own sandbox — exactly the problem Enclave aims to unify.


❌ Common Mistakes

Mistake 1: believing the sandbox replaces code review

An ephemeral container with a network gateway limits the damage; it does not guarantee the quality of the code produced. The agent can perfectly well write mediocre code in its isolated worktree. The solution: the sandbox handles security, human review handles quality. The two don't replace each other.

Mistake 2: allowlisting everything "just to make it work"

The first reflex when facing a blocking gateway: open everything up. That's how you end up with a sandbox that doesn't sandbox anything. The solution: start from an empty allowlist, run the agent, and only add domains as blocks get logged. It's slower the first week, but the resulting policy is real.

Mistake 3: confusing the Docker backend with the microVM

Enclave's default Docker backend isolates well, but the QEMU microVM backend — more airtight — is experimental and has no network restrictions for now. If your threat model requires hardware-level isolation, wait for it to stabilize or stick with strict network policies on Docker.


❓ Frequently Asked Questions

Does Enclave work with all coding agents?

No, but the list is already extensive: Claude Code, Codex, OpenCode, GitHub Copilot CLI, Pi, and Eclipse Theia are natively supported. The architecture is extensible, and a separate repo (enclave-extensions) hosts community harnesses, notably for Antigravity and Playwright MCP. Check the docs for recent additions.

How does it differ from Claude Code's built-in sandboxing?

Claude Code's sandbox (like Anthropic srt) is specific to one editor: the policy lives within the editor's tooling. Enclave externalizes governance: the same policy applies regardless of the agent. If you use multiple agents, it's the difference between five configurations and one.

Is Enclave compliant with the AI Act and the CRA?

Not certified yet. The project's fact sheet indicates that audit trails are aligned with articles 12 and 15 of the AI Act and with the CRA's SBOM requirements, but IT Social points out that Enclave claims no formal compliance. Treat it as a foundation, not a regulatory seal of approval.

Can you run multiple agents in parallel?

Yes, and that's a core use case. Each parallel agent gets its own git worktree, its own containerized session, and its own network gateway. They don't step on each other, and one session going off the rails doesn't drag the others down. That's what makes Enclave a prerequisite for serious multi-agent setups.

How much does Enclave cost?

Free. MIT license, open source, hosted by the Eclipse Foundation. The real cost is the time spent setting up policies and container overhead. Compare that with the time saved: for any team parallelizing agents, the trade-off quickly works in your favor.


✅ Conclusion

Eclipse Enclave 1.0 doesn't reinvent sandboxing — it turns it into a shared, neutral, and auditable infrastructure, where every vendor used to build their own walled garden. With 91% of multi-agent companies considering themselves poorly governed (Deloitte, October 2026), the tool arrives exactly when it's needed. Start with a single agent, a minimalist allowlist, and expand from there — and to choose what you'll run inside it, our comparison of the best AI tools for code is just what you need.