DeepSeek Harness

dsh is a TypeScript monorepo in which everything is a Cordis plugin contributing services, typed events and reversible effects to a shared context. A running instance is a plugin tree composed at boot from a profile: an ordered stack of bundles plus patch files, from a shared base layer (model adapters, tools, persistence, sandbox and approval policy, settings, credentials, telemetry) up to the web, headless, SDK or ACP application. A Python SDK launches the same CLI with the sdk profile.

Multi-agent work runs through a delegation seam. A parent agent can spawn an in-process child or a child in another process; the out-of-process providers include ACP-compatible agents, Codex and Claude Code, and each child is visible to its parent. The Codex and Claude Code providers run one unattended, self-contained text task and return only the final answer, keeping intermediate traffic out of the parent session. Orchestration scripts can fan out children, and background jobs let a parent continue while children run.

State is an append-only session event log with agent-level live events; a goal records durable completion state per session, and scheduled reminders persist across host restarts. The docs describe the loop, tool pipeline and event producers in detail and assume familiarity with Cordis.

The interaction group provides slash commands, one-shot approval prompts that fail closed when no answerer exists, permission presets that combine a sandbox mode (read-only, workspace-write or danger-full-access) with an approval policy, and a tool the model uses to ask the human a question. Plan mode presents a plan for approval before execution. Automation profiles answer approvals through ACP instead of a person, and the Codex and Claude Code child providers run unattended. The safety notice says sandboxing and approvals reduce risk but do not guarantee isolation.

npx @deepseek-ai/dsh web

# Needs Node.js. Starts the Web UI at http://127.0.0.1:3080
npx @deepseek-ai/dsh web
# server only, no browser
npx @deepseek-ai/dsh web --no-open

# From a checkout
git clone https://github.com/deepseek-ai/deepseek-harness.git
cd deepseek-harness
pnpm install
pnpm run build
pnpm dsh web

# Print the plugin tree that a profile boots (with dsh on the PATH)
dsh --profile web --dump-config

• The project is a developer preview and the README warns of compatibility-breaking changes; read SAFETY.md before running it.
• GitHub releases so far are release-candidate pre-releases (for example dsh-v0.2.0-rc.2), not stable versions.
• Run it with least privilege, preferably in a disposable VM or container; sandboxing and approvals do not guarantee isolation.
• The Codex and Claude Code subagent providers use the native authentication and settings of those tools, and run unattended.
• Building from source uses pnpm; pnpm run build must run before pnpm dsh web.
• Desktop and npm CLI share product data but keep packages, activation and lockfiles separate.

Yes, as a client. The mcp-client package connects a configured MCP server and exposes its tools and instructions, and mcp-resources reads its resources. No MCP server mode is documented.

Yes, as subagents. The subagent family includes providers that run a real Codex session through its app-server protocol, a real Claude Code session through the Agent SDK, and any ACP-compatible agent in a subprocess. Each returns a final answer to the parent.

No. The README calls it a developer preview with compatibility-breaking changes expected, and its GitHub releases are release candidates. The safety notice says it has not been security audited.

The repository is developed by DeepSeek AI and includes a DeepSeek API key adapter, but its architecture treats the model adapter as a replaceable plugin. Check the model adapter packages for the providers you need.

The repository is MIT licensed and no paid product is documented. Model API usage is billed by the provider you configure.