Demesne
jbeshir/demesne
Orquestração de agentes IA em containers isolados com controle de segurança
Ver no GitHubMCP server que orquestra agentes de IA e comandos shell em contêineres OpenSandbox descartáveis e isolados. Permite executar scripts não confiáveis, renderizar React, converter mídia e orquestrar múltiplos agentes com segurança via montagens somente leitura e allowlists de egresso.
Instalação
npx convez-ai jbeshir/demesne1 comando no terminal: detecta o formato do repositório (marketplace ou plugin avulso) e instala via Claude Code. Requer Node 18+ e o Claude Code instalado.
Métricas
Casos de uso
- Executar scripts e comandos em sandboxes descartáveis sem risco ao host
- Renderizar widgets React e fazer screenshots via headless browser
- Orquestrar múltiplos agentes IA com tarefas delegadas e verificação
- Transcodificar mídia e processar imagens em ambiente isolado
- Pesquisa autônoma com acesso à internet controlado
Veredito

Ferramenta poderosa para execução segura de código e orquestração de agentes, mas ainda em fase alpha com quebras esperadas. Ideal para quem precisa executar tarefas autônomas complexas mantendo o host isolado por camadas de segurança — contêiner, mount read-only e egress allowlist.
README
A tradução para o português deste README ainda está na fila.
Demesne
An agent-agnostic, local, containerised agent-orchestration MCP server you drive from your agent of choice. It runs untrusted shell, scripts, and AI coding agents in disposable containers, decoupling agent reasoning from execution effects. Host mounts are read-only; outbound network access is governed by egress allowlists.
[!WARNING] Alpha — best-effort. demesne is early software, and is largely built using itself (its own containerised agents do much of the work). Expect rough edges, gaps, and breaking changes between versions. Treat it as alpha and best-effort, and review what it does before relying on it.
What you can do
Ask your agent to run through demesne:
- One-off scripts — execute a shell command in a fresh sandbox and collect output. Example
- Headless React-widget rendering — render and screenshot a React widget inside a sandbox via the demesne-built
browserimage (Playwright + Chromium + Node 22, works ategress=none). How-to - Video / audio / image conversion — transcode, convert, and manipulate media inside a sandbox via the demesne-built
mediaimage (ffmpeg + ImageMagick + libvips + audio tooling, works ategress=none). How-to - Long-running research with open internet — spawn a research agent with unrestricted outbound access. Reference
- Delegated coding-agent tasks — hand off a prompt to a sub-agent running inside a sandbox. Example
- Persistent sessions — create a sandbox, run multiple commands, upload/download files, then destroy it. Example
- Multi-agent orchestration — the orchestrator agent is itself a containerised run that spawns child sandboxes for its workers and verifier, dispatching tasks and judging results across the tree. Example
- Ready-made orchestration skills — a library of
SKILL.mdpipeline definitions (migration sweeps, corpus map-reduce, document ETL, and more) to drop into your agent and adapt. Pre-alpha: in principle ready to use but largely untested — regard them as examples of what could be tried. Example skills
Together these let your agent take on larger tasks more autonomously: you can push security-review-awkward script execution, autonomous research, and entire multi-agent pipelines into containers that run with no permission prompts — much of the autonomy you'd otherwise reach for --dangerously-skip-permissions to get, but with the host kept at arm's length by a container boundary, read-only mounts, and egress allowlists. (That boundary is container-level isolation, not a hard security guarantee — see SECURITY.md.) And you don't pre-declare the pipeline: your agent composes the orchestration prompt itself for the task at hand, and the containerised orchestrator adapts the layout and subagents as it runs.
How it works
Containerised agents can themselves spawn sandboxes, and — with appropriate configuration — get a read-only subset of your host's MCP server tools proxied in through a per-sandbox tunnel. See docs/reference/nested-sandboxes.md.
Get started
a. Install the binary
Download a release binary from the GitHub releases page. Builds are available for linux/amd64, darwin/amd64, darwin/arm64, and windows/amd64.
To build from source instead, see CONTRIBUTING.md.
b. Run a local OpenSandbox
demesne needs a running OpenSandbox instance. See docs/reference/requirements.md for prerequisites; Step 2 of the Quickstart walks through launching one locally.
c. Wire into Claude Code or Codex
See docs/how-to/wire-into-mcp-client.md for the per-client config snippets and full env var reference.
For the full walkthrough, see Quickstart.
Tools
| Tool | Description | Reference |
|---|---|---|
sandbox_script | Run a shell command in a fresh sandbox and tear it down. Returns exit code, stdout, stderr, and the /out host path. | ref |
sandbox_create | Create a persistent sandbox. Returns a sandbox_id handle and the /out host path. TTL is 24h, refreshed by each sandbox_exec. | ref |
sandbox_exec | Run a shell command in an existing sandbox. Refreshes TTL. Returns exit code, stdout, and stderr. | ref |
sandbox_upload | Copy a host file into an existing sandbox. | ref |
sandbox_download | Copy a file out of an existing sandbox; written under <output_dir>/downloads/<basename>. Returns the host path. | ref |
sandbox_destroy | Kill an existing sandbox. Host output dir is preserved. | ref |
sandbox_agent | Run an AI coding agent (codex or claude-code — defaults to codex when Codex credentials are configured, otherwise claude-code) in a fresh sandbox against a caller-supplied prompt. Outbound HTTPS is restricted to the vendor proxy. Returns exit code, stdout, stderr, the /out host path, and the (indicative) cost summary. | ref |
sandbox_research | Run a long-running research agent with no input mounts and unrestricted outbound internet access. Returns exit code, stdout, stderr, the /out host path, and the (indicative) cost summary. | ref |
For a step-by-step walkthrough of the persistent-sandbox lifecycle, see the Quickstart and the sandbox_create / sandbox_exec reference pages.
Docs
| Quickstart | Five steps to your first sandbox_script call |
| Docs | Tutorials, how-to recipes, reference, explanation |
| Examples | Runnable example calls |
| Example skills | Ready-to-use orchestration pipelines you can adapt (pre-alpha) |
Contributing
See CONTRIBUTING.md for building from source, linting, and tests.
Status
See CHANGELOG.md for milestone history.