The "sandbox" topic on GitHub groups 52 open-source projects in the DeepSeek Harness (DSH) ecosystem, led by layerfs with 2.7k GitHub stars. layerfs — Ephemeral Workspaces, durable shared history. Every project here is indexed by DSH Universe with live GitHub data — stars, activity and install status — so you can compare and install directly.
Ephemeral Workspaces, durable shared history. LayerFS gives every agent a complete, writable filesystem Workspace over shared committed history, without copying the base. Agents work in parallel, commit what changed as deduplicated checkpoints, then branch, rewind or merge. Try it with LayerFS × DSH.
Local-first AI agent runtime with sandboxed sessions, MCP tools, memory, credentials, audit/replay, and a built-in console. Run OpenAI, Anthropic, MiniMax, DeepSeek V4, and OpenAI-compatible models on your infrastructure.
Second-model AI auto-review for DeepSeek Harness approval requests: a read-only reviewer subagent returns structured allow/deny verdicts with reasons, fail-closed by default, fully auditable from the session log (approval/asked -> autoReview/verdict -> approval/decided).
Minimal surface, complete architecture: a small, local-first coding-agent harness that re-derives the invariants of DeepSeek Harness with fewer concepts. Independent educational project, not affiliated with DeepSeek.
Prevents the model from rewriting specified paths in the workspace. Also lets you open up writes to certain external directories without opening the sandbox.
DeepSeek Harness auto-approval plugin: deterministic rules + LLM safety review layered on top of the workspace-write sandbox; auto mode never loosens the sandbox, fail-closed. Safe auto-approval for DeepSeek Harness — deterministic rules + LLM review on top of the workspace-write sandbox. Auto mode without ever granting full-access.
DSH plugin developer sandbox: spawn isolated DeepSeek Harness web mirror instances (own DSH_HOME/port/profile) that mount the plugin under development, with optional host API/model inheritance. GUI panel + sandbox_* agent tools.
Codex-style automated approval review for DeepSeek Harness: an isolated reviewer subagent decides sandbox escalations by intrinsic risk and user authorization. Community Beta — not an official DeepSeek AI plugin.
Rayrun cloud computers for DeepSeek Harness: create, manage and open remote Linux desktops (noVNC) in the sidebar, and hand agent tasks to the same machines with rayrun_* tools. Per-session cloud/local routing, password or WeChat login, token kept in the Harness vault.
Run DeepSeek Harness against a gVisor-isolated cloud sandbox — files, Bash, PTY and LSP all move off your machine. Drop-in plugin, no fork, optional persistence.
Crash-resilient remote execution world for DeepSeek Harness: ctx.fs/ctx.subprocess over an E2B sandbox with heartbeat keep-alive, transparent recovery, and workspace sync.
DSH Web GUI plugin: session-scoped "Allow for this session" option in the approval dialog (per-mode standing grants, localStorage) + right-panel auto-open guard
Claude Code-style declarative permission rules for DeepSeek Harness: ordered allow/deny/ask rules with tool-name, argument (glob/regex), and workspace-path matching on the tools/pre-execute waterfall, session-log audit, and HMR reload.
Sandboxed install validation for DeepSeek Harness plugins — installs each one into a fresh profile in a throwaway container and reports what the harness recorded
Per-command sandbox allow-list for the DSH bash tool: allow-listed CLIs (lark-cli, gh, ...) run under a home-writable sandbox instead of prompting for danger-full-access.
DeepSeek prices, plugin search and model specs inside DeepSeek Harness. We run free sandbox install tests on every DSH plugin — see which ones actually work.
Claude Code-style declarative permission rules for DeepSeek Harness: ordered allow/deny/ask rules with tool-name, argument (glob/regex), and workspace-path matching on the tools/pre-execute waterfall, session-log audit, and HMR reload.
Claude Code-style declarative permission rules for DeepSeek Harness: ordered allow/deny/ask rules with tool-name, argument (glob/regex), and workspace-path matching on the tools/pre-execute waterfall, session-log audit, and HMR reload.
Claude Code-style declarative permission rules for DeepSeek Harness: ordered allow/deny/ask rules with tool-name, argument (glob/regex), and workspace-path matching on the tools/pre-execute waterfall, session-log audit, and HMR reload.
Container-isolated backend for the DeepSeek Harness code-execution seam: Code Mode programs run in a fresh container with no network, a read-only rootfs, and kernel-enforced memory, CPU, and pid ceilings
Claude Code-style declarative permission rules for DeepSeek Harness: ordered allow/deny/ask rules with tool-name, argument (glob/regex), and workspace-path matching on the tools/pre-execute waterfall, session-log audit, and HMR reload.
⚠️ NOT REVIEWED BY ANY HUMAN - AI-generated (DeepSeek V4.1-flash). DeepSeek Harness desktop 0.2.0-rc.2 (Windows) bundle: gives the ACL sandbox runner a console so confined pwsh/cmd/powershell start again under workspace-write. No confinement is relaxed. dsh-plugin.
Permission matrix for DeepSeek Harness: 3 sandbox modes x 4 approval strategies = 9 presets, with global and LLM-robot defaults, three-tier risk policies, audit log and git checkpoint.
Remote-only OpenKapsel workspace bridge for the DeepSeek Harness: replaces host filesystem and shell tools with a fail-closed remote tool set. Self-hosted alternative to cloud agent sandboxes.