LATEST
AI Tools

Eclipse Enclave 1.0: A Free Sandbox for AI Coding Agents

Updated November 19, 2026
15 min read
Editor's Note: This article was human-reviewed for accuracy
Close-up of colorful programming code on a dark computer monitor

Quick Answer & Key Takeaways

  • • Eclipse Enclave 1.0, launched on November 19, 2026, is a free, open-source sandbox that runs AI coding agents in isolated containers.
  • • It works with Claude Code, Codex, OpenCode, Copilot and more, and it limits each agent to the files and websites you allow, with a log of everything it reaches.
  • • Enclave itself costs nothing under the MIT license; the only bill is the coding agent you choose to run inside it.

AI coding agents are at their most useful when nobody is watching them. Give Claude Code or Codex a task and let it run, and it will write the code, install the packages, run the tests and fix what breaks, all without asking you to approve every step. That same freedom is the problem. An agent running loose on your machine can delete the wrong files, paste a secret into the wrong place, or follow a poisoned instruction hidden in a web page it read. Eclipse Enclave 1.0, launched on November 19, 2026, is the Eclipse Foundation's answer: a free, open-source sandbox that gives an agent full autonomy inside a container, and nothing outside it.

Enclave is not another coding agent. It is the room the agent works in. You point it at a project, it builds an isolated container around that project, and your agent of choice runs inside it, cut off from the rest of your machine. The project is small, it is day one, and it is honest about that. What makes it worth your attention is who ships it, what it costs, and the specific problem it solves for anyone already letting agents touch real code.

What Eclipse Enclave 1.0 actually is

Eclipse Enclave 1.0 is an open-source, vendor-neutral runtime and sandbox for AI coding agents, released under the MIT license by the Eclipse Foundation on November 19, 2026. The Foundation is one of the largest open-source software foundations in the world, home to more than 450 projects including the Eclipse IDE, Jakarta EE and the Open VSX Registry, and it is backed by over 300 member organisations. Enclave is part of its growing AI ecosystem, alongside the Theia AI-native IDE, Eclipse LMOS for multi-agent systems and the Sovereign AI Foundation.

The project started at EclipseSource, an Eclipse Foundation member organisation, where engineering teams built sandboxing technology so their own developers could let coding agents work autonomously without exposing their machines. EclipseSource contributed the technology to the Foundation to make it available to a wider community. Dr. Philip Langer, the project co-lead and CEO of EclipseSource, puts the trade-off plainly: agents deliver real productivity gains only when developers can give them freedom to work independently, and Enclave exists to make that possible without handing agents unrestricted access to development machines.

The timing is not accidental. Coding agents have moved past suggestions. They now write and modify software, execute commands, install dependencies and run tests on their own. The Eclipse Foundation's announcement says that realising their full productivity potential requires giving agents greater autonomy, and that running them with unrestricted access to development environments can expose sensitive information, cause unintended changes and create governance problems for organisations trying to scale adoption. Enclave is the Foundation's bet that the fix is isolation with control, not more confirmation prompts.

An industry analyst report from 451 Research, part of S&P Global, cited in the launch announcement, identifies Enclave's consistent isolation and policy controls across supported coding agents as a source of immediate enterprise value. That is the vendor's own announcement citing the report, so treat it as the Foundation's framing rather than independent verification. No independent reviews of Enclave exist yet. It launched today.

How the sandbox works

Enclave puts four controls around each agent session, and the docs describe them as See, Run, Reach and Keep. Each one answers a different way an agent can hurt you.

See: the agent only sees what you mount. Each session runs in its own container with its own filesystem, process tree and network stack. Only the paths you mount, usually the project worktree, are visible to the agent. Everything else on your machine stays outside the container. You can mount extra directories, such as documentation or a related project, when a task needs them, and the configuration supports per-project and per-tool overrides.

Run: an ephemeral container per session. The container is created for the session and thrown away when it ends. Parallel sessions never share a container, and Enclave gives each one its own git worktree, using plain git with no extra setup, so two agents working at once cannot collide with each other's changes. You review and merge on your own terms. The worktree approach is simple and it is the part of the design most likely to survive contact with real teams, because it uses a mechanism developers already trust.

Reach: only allowlisted domains, through a gateway. With the default Docker backend, a sidecar gateway filters DNS and proxies outbound traffic. The agent reaches only the domains on your allowlist. Everything else is blocked and logged. This is the control that matters most against prompt injection, because a hijacked agent that cannot reach an attacker's server cannot exfiltrate much. Secrets configured for HTTP release are masked from the agent: the gateway injects them only into requests to matching allowlisted hosts, so their values stay hidden. Other credentials arrive inside the container as mounted files or environment variables.

Keep: sessions persist. You can start, pause, resume and inspect a session whenever you want. Authentication, history and configuration persist across restarts, including custom skills and MCP servers. The gateway logs every connection it proxies and every lookup it blocks, which gives you an evidence trail for security and compliance reviews. You can audit the log with enclave network log.

The docs also sketch where this is heading for regulated teams. The same foundation, container isolation, network allowlists and gateway logs, is meant to grow toward requirements such as the EU Cyber Resilience Act, with SBOM evidence and vulnerability reporting, and the EU AI Act, with record-keeping and resilience to unauthorised use. Those are stated goals in the documentation, not shipped features, so do not buy Enclave today for a compliance deadline tomorrow.

Which agents it runs

Enclave is deliberately agent-neutral. It currently supports Claude Code, Codex, Copilot, OpenCode, Pi and Theia AI, with an extensible architecture for more. Tools that are not built in, such as Amp or Antigravity, live in the official enclave-extensions repository and can be added with enclave tools add. The point is that you pick the agent and Enclave provides the same isolation around all of them, so switching agents does not mean rebuilding your safety setup.

AgentVendorBuilt inHow you pay for it
Claude CodeAnthropicYesClaude Pro $20 a month, or $17 a month on annual billing; Max $100 a month
CodexOpenAIYesChatGPT Plus $20 a month, with limited Codex access on the free plan
CopilotMicrosoft / GitHubYesThrough your GitHub Copilot plan; billed by GitHub, not by Enclave
OpenCodeOpen sourceYesFree and open source; you bring your own model API key
PiOpen sourceYesFree and open source; you bring your own model API key
Theia AIEclipse FoundationYesFree and open source
Amp, AntigravityVariousVia extensions repoThrough each tool's own plan

If you already use Claude Code, our Claude Code explainer covers what the tool does in the terminal and what its plans include. If you want the free option, our OpenCode review covers the open-source agent you run on your own machine, which pairs naturally with Enclave because both cost nothing.

Installing it and starting your first session

Enclave needs three things before it will do anything: a supported host, a container backend and Git. Supported hosts are Linux, macOS, or Windows through WSL2. The container backend is Docker by default. Git is required because Enclave runs from inside a git repository and uses worktrees to isolate parallel sessions. You also need credentials for whichever agent you plan to run.

Installation is a download, not a package-manager ritual. Prebuilt binaries, a Debian package and an RPM package are published with each release on GitHub. On Ubuntu 24.04 you install the .deb with APT so the runtime dependencies resolve. On Fedora you install the .rpm with DNF. On other Linux distributions you download the binary for your architecture, verify it against the published checksums and install it into ~/.local/bin. On macOS you do the same with the Apple Silicon or Intel binary. On Windows there is no native build: you install WSL2 with Ubuntu 24.04, make Docker available inside it, and follow the Linux instructions, with an optional Windows launcher installed through Scoop that forwards commands into WSL. You can also build from source with the Go toolchain by cloning the repository and running make build.

Two rough edges are worth knowing before you start. The macOS binaries are unsigned and not notarised, so a download through the browser carries a quarantine attribute that makes Gatekeeper refuse to run it. The docs tell you to strip it with xattr first, or download with curl or gh release download to avoid the attribute entirely. And the docs themselves carry a note that they are still being filled in, with the commands and paths reflecting the current CLI. This is a 1.0 with fresh documentation, not a polished manual.

Once installed, the first session is one command from inside a git repository:

enclave

The first interactive run asks which agent to use and saves the answer, so you are only asked once. Enclave builds the environment, mounts the current folder into a container and starts your agent against the branch you have checked out. The agent runs at full autonomy with no confirmation prompts, and it stays contained.

Three more commands cover daily use. enclave --tool codex picks a different agent for one run. enclave continue resumes a previous session. enclave ps lists active sessions. For parallel work, create a git worktree per task, start Enclave from inside each one, and merge when you are ready. A graphical companion called HomeShell is in preview: it groups work into projects and workstreams and lets you start, watch and answer agent sessions from one web UI or desktop window. It has no stable release yet, and its builds follow every change, so treat it as an early preview rather than a finished product.

What it costs

Enclave itself is free. It is MIT-licensed open-source software, there is no paid tier, no hosted service and no per-seat price. The only money involved is whatever the agent inside it costs, plus the machine you already own. That is the whole pricing story, and it is the sharpest contrast with hosted sandboxes.

What you pay forUSDGBP (approx.)CAD (approx.)Notes
Eclipse Enclave 1.0$0£0CA$0MIT license, no paid tier
Claude Code, via Claude Pro$20 a month, or $17 a month billed annually£15.20, or £12.90 annuallyCA$28.60, or CA$24.30 annuallyClaude Code included in all paid Claude plans; Max is $100 a month
Codex, via ChatGPT Plus$20 a month£15.20CA$28.60Limited Codex access on the free plan; no annual billing for Plus
OpenCode$0£0CA$0Free and open source; you pay your model provider's API rates

GBP and CAD figures are approximate, converted at November 19, 2026 market rates of 1 USD to about £0.76 and about CA$1.43. Your card statement will differ slightly with fees and your provider's exchange rate. Prices shown do not include applicable tax. The Claude figures come from Anthropic's own pricing page, and the ChatGPT figures from OpenAI's pricing page, both checked on November 19, 2026.

The honest total for a typical setup is therefore $20 a month for one agent subscription, or $0 if you pair Enclave with OpenCode and a pay-as-you-go API key. Either way, the sandbox adds nothing to the bill. That matters because the alternative most teams reach for, a hosted sandbox, charges for compute by the hour even when the agent is idle waiting on a model call.

Enclave vs the alternatives

There are three ways to run a coding agent today, and Enclave is the middle path. Running the agent bare on your host costs nothing and gives it everything, including your SSH keys, your browser cookies and your other projects. That is how most developers started, and it is the setup Enclave exists to replace. Running it in a hosted sandbox, such as E2B or Daytona, gives you strong isolation in someone else's cloud, billed by usage. Enclave gives you the isolation on your own machine, for free, with the trade-off that you run and maintain it yourself.

Agent on your hostEclipse Enclave 1.0Hosted sandbox (E2B, Daytona)
Cost$0$0About $0.17 an hour for a 2 vCPU, 4 GB sandbox; E2B Pro is $150 a month
Where code runsYour machine, unrestrictedYour machine, in a containerThe provider's cloud
Network controlNoneDomain allowlist plus gateway logsVaries by provider and tier
Parallel agentsThey share your filesSeparate git worktree eachSeparate sandbox each
Source codeDepends on the agentOpen source, MITE2B open source; Daytona closed source since June 2026
Setup effortNoneInstall Docker and EnclaveSign up and call an API

The hosted-sandbox rates come from third-party pricing comparisons checked in October and November 2026, which put both E2B and Daytona at $0.0504 per vCPU-hour and $0.0162 per GiB-hour. A 2 vCPU, 4 GB sandbox therefore costs about $0.17 an hour while it is alive, and agents spend most of their time idle waiting on model calls. E2B's Pro plan adds $150 a month for longer sessions and higher concurrency. None of that makes hosted sandboxes a bad deal. They are the right choice when you need to run agents for other people, or at a scale no laptop can hold. Enclave is the right choice when the agent works for you, on your machine, and you would rather pay nothing than rent a computer you already own.

Our kagent 1.0 article covers the other end of the same problem, running agents on Kubernetes without a pod per agent, for teams that need orchestration rather than a single developer's sandbox.

What it cannot do yet

A 1.0 launched today deserves a hard look at its limits, and Enclave has several that the docs state openly.

First, it is small. The GitHub repository, eclipse-enclave/enclave, was created on July 9, 2026, and at launch it has 86 stars and 18 forks. That is not a community yet, it is a starting line. Compare that with the hundreds of thousands of stars on the projects it wraps, and the gap is the point: Enclave is unproven in the wild, with 46 open issues at launch and no track record of security fixes under pressure.

Second, the experimental QEMU microVM backend currently runs without network restrictions. If you choose it for stronger isolation, you lose the gateway filtering that is the whole reason to use Enclave, until that gap is closed. It is also x86-64 only and practical only on x86-64 Linux hosts where KVM can accelerate it. The Docker backend is the one to use today.

Third, a container is isolation, not a guarantee. An agent inside the sandbox can still misuse anything you mounted or any domain you allowlisted. If you mount your whole home directory and allowlist the open internet, you have rebuilt the original problem inside a box. The sandbox only helps if the allowlist is short and the mounts are narrow. Prompt injection is still possible inside the container; what Enclave limits is what an injected agent can reach, not whether it can be injected.

Fourth, the rough edges are real. The macOS binaries are unsigned, HomeShell has no stable release, and the docs are still being filled in. There is no Windows native build, only WSL2. None of this is a reason to skip Enclave, but it is a reason to treat the first week as an experiment on a side project, not a rollout across your team.

Common mistakes to avoid

Mounting too much. The default mount is the project worktree. Adding your home directory, your credentials folder or another project because it is convenient hands the agent the keys you were trying to protect. Mount only what the task needs.

Allowlists that allow everything. A network allowlist with a wildcard or with every registry and package host your agent might ever want is not an allowlist. Start narrow, add domains when a build fails, and read the gateway log to see what was blocked.

Running parallel agents on the same checkout. Two agents editing the same files will collide. Use a git worktree per session. It is plain git, it takes one command, and it is the difference between parallel work and a merge conflict.

Using the QEMU backend for untrusted code. The experimental microVM backend has no network restrictions today. Until that changes, untrusted work belongs on the Docker backend with the gateway in front of it.

Treating the sandbox as a substitute for reviewing the diff. Enclave limits what an agent can touch, not whether its changes are correct. Read the diff before you merge. The sandbox protects your machine; only you protect your codebase.

Frequently asked questions

Is Eclipse Enclave really free?

Yes. Enclave 1.0 is open-source software under the MIT license, with no paid tier and no hosted service to subscribe to. You download it, run it on your own machine, and pay nothing. The only cost is the coding agent you run inside it, such as Claude Pro at $20 a month or ChatGPT Plus at $20 a month, or nothing at all if you use a free open-source agent such as OpenCode with a pay-as-you-go API key.

Which AI coding agents work with Enclave?

Claude Code, Codex, Copilot, OpenCode, Pi and Theia AI are built in, and more can be added through the official enclave-extensions repository, which already covers tools such as Amp and Antigravity. You pick the agent with a flag such as enclave --tool codex, and Enclave provides the same isolation around each one.

Do I need Docker to use Enclave?

Yes, for the default backend. Enclave runs each agent session in a container, and Docker is the container backend it uses today. An experimental QEMU microVM backend exists, but it currently runs without network restrictions, so the Docker backend is the one to use. You also need Git, because Enclave runs from inside a git repository.

Does Enclave stop prompt injection?

No, and it does not claim to. Prompt injection happens inside the agent, and Enclave cannot stop an agent from following a malicious instruction hidden in text it reads. What Enclave limits is the damage: the injected agent can only touch the files you mounted and reach the domains on your allowlist, and every connection it makes is logged. Keep mounts narrow and allowlists short, or the sandbox does little.

Should I use Enclave or a hosted sandbox such as E2B or Daytona?

Use Enclave when the agent works for you on your own machine and you want isolation without a bill. Use a hosted sandbox when you need to run agents for other people, at a scale beyond one machine, or through an API your product calls. Hosted sandboxes charge roughly $0.17 an hour for a typical 2 vCPU, 4 GB sandbox, so a solo developer running a few sessions a week will usually spend less, which is to say nothing, with Enclave.

The bottom line

Eclipse Enclave 1.0 solves a real problem with a simple idea: let the agent work alone, in a room of its own, with a door you control. It is free, open source, vendor-neutral, and it works with the agents developers already use. It is also day one, with a small community, fresh documentation and a few sharp edges. Try it on a side project this week, keep the mounts narrow and the allowlist short, and read the diff before you merge. If it survives that test, it is one of the lowest-cost security upgrades your coding agent will ever get.

Sources

Exchange rates: November 19, 2026 market rates, verified at publication (1 USD ≈ £0.76, ≈ CA$1.43). GBP and CAD figures are approximate.

15 min read
157 views
edit_note Written by

Ratul Ahamed

Founder & Editor-in-Chief

Ratul Ahamed is the founder and editor of AI Core Updates, an independent site covering AI tools, models and pricing. He is a web developer based in Bangladesh and also runs All In One Tools, a free collection of browser-based utilities. Contact: [email protected]

Read all stories by Ratul Ahamed arrow_forward
verified_user Human-Reviewed Editorial

forum Comments (0)

Leave a Reply

chat_bubble_outline

No comments yet. Be the first to join the conversation!

Weekly AI Digest

Get our weekly AI tools digest

One email a week analyzing new AI tools, model updates, and practical workflows. Tested and summarized with zero fluff. Unsubscribe any time.

bolt
Hands-on AI Tool Reviews
Pricing math, limits, and alternatives.
verified_user
Strict Anti-Spam Policy
1-click unsubscribe at any time.
mark_email_unread

Subscribe to Digest

Sent weekly • No spam

person_outline
mail_outline

No spam ever. You can unsubscribe with a single click at any time.

cookie

Cookie Settings

We use cookies to enhance your browsing experience, serve personalized content, and analyze our traffic. By clicking "Accept All", you consent to our use of cookies. Learn more