DocsProviders
DeepSeek Harness
DeepSeek Harness
(dsh) is DeepSeek's open-source, plugin-based agent harness. Moonfold
runs it through its ACP profile (dsh --profile acp), with the model providers
you configured in your own DeepSeek Harness setup.
Add it
- Open Settings → Providers → Add provider.
-
Under Or choose from ACP Registry, search for
DeepSeek Harness and select Add. Moonfold installs the
pinned
@deepseek-ai/dshpackage withnpxon the machine that runs the Moonfold server. Node.js must be installed there. - Skip the sign-in step: DeepSeek Harness has no in-app sign-in. It uses the providers and keys in its own configuration.
Give it a model
DeepSeek Harness keeps its configuration in ~/.dsh (or
$DSH_HOME). The easiest way to add a model is its own web UI: run
npx @deepseek-ai/dsh web once, open Settings → Models, and
add your DeepSeek API key or another provider. You can also export
DEEPSEEK_API_KEY in the instance's Environment in
Moonfold's provider settings.
To use a local model, add a custom provider in DeepSeek Harness that points at Ollama
(http://127.0.0.1:11434/v1) or LM Studio
(http://127.0.0.1:1234/v1) with the OpenAI Chat Completions protocol. Its
models then appear in Moonfold's model picker under DeepSeek Harness.
The ACP profile has its own default model. Pick the model you want in the composer: Moonfold sends your choice, and the reasoning effort, to DeepSeek Harness at the start of each turn.
What works
- Threads, steering by sending another message, stopping a turn, model and reasoning-effort pickers, Moonfold's MCP tools and your MCP servers.
- Handoff: when DeepSeek's API reports an exhausted balance or quota, or a rate limit, the turn ends as a usage limit and Moonfold offers to continue in another agent.
- Not available through ACP: DeepSeek Harness's own plans, todos, terminal views and questions; it shows its work as messages and tool calls instead.
Moonfold is not affiliated with or endorsed by DeepSeek.