Sandbox & Runtime
Runtime, sandboxing, execution, and environment helpers.
HOL Guard is an open-source antivirus for AI agents that acts as a local-first security layer for agents, tools, plugins, skills, MCP servers, and package installs. It reviews actions before they run, detects secret exposure, destructive operations, prompt injection, and supply-chain risks, and allows, blocks, or requests approval based on policy.
基于上游 1.1.71 维护的 dsh-purge 社区兼容包:只负责 DeepSeek Harness 的版本与本地 Windows 运行问题兼容。
Unofficial DeepSeek Harness plugin for rule-gated automatic approval of Shell and PowerShell sandbox escalations. It applies fixed high-risk checks and literal command-prefix rules, optionally uses a tool-free LLM reviewer, grants only one allowed-once per decision, and returns uncertain requests to native human approval.
Zero-dependency Node gateway for multi-user DeepSeek Harness server deployments, providing a login portal, per-user DSH instances isolated by OS accounts, per-user API keys, and a delivery file drawer. It runs on a remote server and is not a local desktop tool.
A multi-tenant DeepSeek Harness (DSH) hosting platform intended for deployment on a public server. Users register and are approved by an administrator, then receive isolated DSH environments accessible through domain-based URLs.
QAQ is a launch resilience guard for DeepSeek Harness (DSH) by WTStarMark. It supervises the `dsh web` process, reads the browser's real DOM through headless Chrome + CDP, detects host crashes, Web UI red screens and plugin degradation, then restores the configuration snapshot from the last successful boot and restarts, preserving the broken config for manual recovery. It is described as non-invasive, shipped as a one-command script with structured logs, and includes a plugin hot-update guard (c
A Windows plugin for DeepSeek Harness that adds optional Git Bash, WSL Bash, MSYS2 Bash, and SSH Bash tools plus persistent PTY terminal sessions. git_bash and terminal are enabled by default; pwsh remains served by the base bundle.
Unofficial community plugin for DeepSeek Harness web that adds a middle approval tier between workspace-write and danger-full-access: safe operations in trusted directories are auto-approved, while destructive operations always prompt for human confirmation. Includes Standard/Aggressive modes, 11 per-category switches, an in-plugin audit trail, and a persisted config at ~/.dsh/perm-guard.json.
Docker packaging of the npm-released DeepSeek Harness (DSH), published for linux/amd64 and linux/arm64. It preinstalls DSH and its runtime dependencies so the host needs no Node.js or npm packages, supports LAN WebUI access, offers an optional shared admin key, runs as a non-root user with recommended read-only root filesystem and dropped Linux capabilities, and keeps configuration and workspace data persistent. The README warns against exposing the service directly to the public internet.
dsh-safety is a filesystem safety harness plugin for DeepSeek Harness. It enforces a three-tier file policy at tool execution, routes deletes through trash, requires human approval for sensitive deletes/writes, snapshots and rolls back plugin composition, and validates composition before restart. It includes model-facing safety tools, an audit journal, and a standalone zero-dependency CLI.
W8 sandbox plugin for DeepSeek Harness
给 DeepSeek Harness 加上下一句 Tab 补全与通俗摘要浮窗。Concise next-message suggestions and editable summaries. MIT.
面向 DeepSeek Harness 的《战争雷霆》自定义涂装插件:四边形拾取器、声明式图层、支持 4K DDS 纹理与.blk 文件导出。本项目由DeepSeek在DSH上完成。War Thunder user-skin studio for DSH: quad picker, declarative layers, 4K DDS + .blk export.
让文本模型读懂一首歌:本地音乐分析插件,把音频拆成能量、速度、调性、密度,音频不出本机。
An out-of-repo DeepSeek Harness plugin that applies a host allowlist to the web_fetch and web_search tools at connect time, and reads the destinations named in bash, pwsh or run_code arguments before the command runs. It ships in audit mode, records one OCSF Network Activity (4001) record per decision, and includes a standalone report CLI.
Permission gate and audit plugin for DeepSeek Harness. It intercepts tools with layered permissions, routes to one-time approval when needed, and records audit entries in ~/.dsh/agent-gate/audit.jsonl.
A DeepSeek Harness plugin that adds a Permissions tab to any session and computes effective permissions as the stricter intersection of the official preset, the agent policy, and per-session overrides. It also pins the official sandbox/mode for Claw sessions (tightening only).
DeepSeek Harness plugin for GPT/Codex sandbox escalation parameter compatibility. It removes redundant sandbox_permissions and justification fields only when the request can be proven redundant, while preserving actual escalation requests for DSH native approval.
dsh-worlds is a DeepSeek Harness plugin that runs the agent's execution world inside a Docker container, so Bash, persistent PTY terminals, LSP, and file tools can stay in a container that outlives the harness process. It uses only node:http, node:net, and node:crypto with zero npm dependencies, and is not affiliated with DeepSeek AI.
A plugin for the DeepSeek Harness (dsh) that provides an unattended full-access (yolo) window: a human-only /yolo command arms danger-full-access sandbox plus never approval for that session, with lazy timed expiry and automatic revert, a best-effort catastrophic-command tripwire guard, and desktop/webhook/email notifications. The README states the bundled guard is a best-effort tripwire against accidents, not a security boundary.
Run OmicVerse/OmicOS bioinformatics analyses from DeepSeek Harness via a persistent Python kernel, with capability catalog search and account tab.
A zero-dependency DeepSeek Harness plugin that acquires an exclusive lock on the DSH_HOME data directory at startup, aborting loudly if another live dsh server is already using it, to prevent session-log corruption from concurrent writers.
Linux backend for dsh-witness/dsh-anchor: sandbox recipes (chattr + bubblewrap) and process identity tools. Every capability claim is backed by experiment numbers and control groups.