DeepSeek Harness with Claude Code: Integrate or Migrate
Use DeepSeek Harness alongside or instead of Claude Code. ACP integration, the subagent pattern, and what a migration looks like in August 2026.
Two Ways to Combine Them
DeepSeek's own positioning is that dsh is 'less a competitor, more a framework that can sit above them' — you can call Claude Code or Codex as a subagent from inside dsh[1]. That opens two practical setups: run both in parallel, or migrate fully.
The parallel setup keeps your existing Claude Code muscle memory and adds dsh where its strengths pay off: model freedom, full traceability, and plugin-level customization. The migration path moves your workflows to dsh and closes the Claude subscription[2].
Either way, the mechanics are the same: both tools speak the same kinds of protocols, and dsh's ACP/MCP channels make the handoff explicit.
Keep Claude Code, Add dsh as a Subagent
dsh can treat Claude Code (and Codex) as callable subagents, so you delegate the model-specific work while dsh owns orchestration[1].
This is the 'sit above them' architecture: dsh is the chassis, Claude Code is one of the tools plugged in. Teams that cannot drop Anthropic for certain workloads (long-form reasoning, specific tooling) use this hybrid to keep one runtime.
- dsh orchestrates; Claude Code executes model-specific tasks as a subagent
- You keep Claude Code's strengths without abandoning the harness
- The subagent pattern is a first-class plugin capability, not a hack
ACP Channel: Drive dsh from Claude-Style Clients
If you want the reverse — drive dsh from a client you already use — the community ACP plugin exposes dsh as an Agent Client Protocol server, and ACP-capable editors (Cursor, Zed, JetBrains) connect to it[3].
Model choice inside dsh is a config field, not a fork: the same profile can route to DeepSeek V4, Claude, GPT, or Gemini through providers — see the model configuration guide.
# Install the ACP channel and start the stdio server
dsh plugin --profile acp add @jeremy9682/dsh-acp
dsh --profile acp
# Point your ACP client at the server; dsh now appears
# alongside or instead of Claude Code in the same editorFull Migration Checklist
For teams moving from Claude Code to dsh, here is the practical order[2][4].
- Install dsh and open the Web UI (npx @deepseek-ai/dsh web)
- Add your model provider (DeepSeek first; add others later)
- Port MCP servers — dsh uses MCP via plugins (see MCP setup)
- Recreate Agent Skills as dsh skills plugins
- Re-map your keybindings/scripts to dsh --profile headless
- Run two weeks of parallel A/B on real repos before dropping the subscription
Remember dsh is a developer preview with breaking changes ahead — keep a Claude Code fallback until your critical workflows pass on dsh. The compatibility risk is real in August 2026[2].
What You Give Up
A fair migration comparison lists what you lose as well as what you gain.
The 262 MiB story matters here too: early dsh rcs shipped bundled Claude Code components that were later removed. If you find Claude Code internals inside a dsh install, you are on an old rc — upgrade[2].
| You gain with dsh | You give up vs Claude Code |
|---|---|
| MIT open source, self-hostable, auditable loop | First-party Anthropic support and roadmap |
| Any model, any provider, no subscription | Polished managed MCP/plugin market |
| Pluggable everything (loop, UI, sandbox) | Known-good default agent behavior |
| Append-only session log + Trajectory view | Claude Code's mature CLI ergonomics |