Run Harness on your own hardware: the DeepSeek Harness plugin
The dsh-agent-harness plugin routes bounded coding tasks to local or cloud models through DeepSeek Harness, with the same independent verification either way. Here's what that buys you.
By Agent Software
Local models were already supported. This is different.
Agent Harness has always been able to route tasks to local models through its provider router. The DeepSeek Harness plugin — dsh-agent-harness — is a different kind of integration: it connects Agent Harness directly to DeepSeek Harness (DSH) as its own worker lane, so DSH tasks show up in the same pool as everything else, run against local models via LM Studio or against cloud models, and get verified exactly the same way regardless of which one you picked.
harness plugin add dsh-agent-harness
harness worker add dsh \
--endpoint http://localhost:1234/v1 \
--model qwen3-coder
harness run tasks/ts-lib/ts-isUUID.json \
--worker dsh
The trust model doesn't change
The easy way to build a local-model integration is to quietly relax verification for it — local is cheap, so who cares if it's a little less rigorous. We didn't do that. A DSH worker's claim of "done" gets the same independent check a cloud worker's claim gets. Doctor checks, structured reports, and audit logs all apply the same way. You're not trading verification for cost savings; you get both.
Why this is worth doing
The direct benefit is obvious: offload coding tasks to cheaper or free local models running on your own machine, and cloud API tokens stop burning on work your hardware can already handle. For a lot of bounded, well-specified tasks — the kind Agent Harness is built around — a local model is plenty, and every one of those tasks you route locally is a task you didn't pay a cloud provider for.
The less obvious benefit is control. Local-or-cloud, per task, without changing how you call Harness, means you get to make that cost/quality tradeoff yourself instead of having it made for you by whichever provider you happened to configure first.
We've been running the DSH plugin against real bounded task suites as part of ongoing internal testing. It's not a demo bolted on for a launch post — it's a live integration we keep extending as local models get better, which they keep doing.
Take control of your AI workflow today.
