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.
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 - Offline interactive-fiction build/playtest — compile a Twine story and playtest it headlessly via the demesne-built
twineimage (Tweego + Twine story formats + Chromium, works ategress=none). Reference - Offline HTML5-game build/playtest — build and playtest a web game via the demesne-built
webgamedevimage (a warm Phaser + Vite + TypeScript template + Chromium, works ategress=none). Reference - 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.
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.
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.
demesne needs a running OpenSandbox instance. See docs/reference/requirements.md for prerequisites; Step 2 of the Quickstart walks through launching one locally.
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.
| 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 48h, 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 in a fresh sandbox against a caller-supplied prompt. Provider is inferred from model (defaults to the first enabled provider with configured credentials, Codex-first). 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 |
All three spawn tools (sandbox_script, sandbox_agent, sandbox_research) accept an optional background: true parameter. When set, the tool returns immediately with {job_id, status: "running"} instead of blocking. Use the complementary job-control tools to manage the run:
| Tool | Description |
|---|---|
sandbox_status |
Non-blocking snapshot of status, elapsed time, and cost; pass include_stdout_tail: true for the bounded stdout tail. |
sandbox_wait |
Block up to timeout_seconds (default 1800, max 172800) for a terminal state; returns the final result or a "still running" sentinel if the timeout elapses. |
sandbox_cancel |
Cancel the job and its entire descendant subtree; idempotent on already-terminal jobs. |
Synchronous calls may run to completion (subject to the explicit 48h runtime limit) and remain cancellable. Use background: true for concurrent work, detachment, status/progress polling, or deliberate job control. A background run automatically attempts one advisory MCP notifications/message logging notification when it becomes succeeded, failed, or cancelled; synchronous runs do not. No extra opt-in beyond background: true is required. Client display, wake, or model-context injection is not guaranteed, so sandbox_status / sandbox_wait remains the reliable fallback. See background notification compatibility. sandbox_wait is a bounded wait (30-minute default, 48-hour maximum), not a runtime limit. The job registry is in-memory; jobs do NOT survive MCP-server restarts (a stale job_id returns an error after restart); completed jobs are retained ~1h via a TTL reaper.
For a step-by-step walkthrough of the persistent-sandbox lifecycle, see the Quickstart and the sandbox_create / sandbox_exec reference pages.
| 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) |
See CONTRIBUTING.md for building from source, linting, and tests.
See CHANGELOG.md for milestone history.