实测 DeepSeek Harness 一周：产出一个已发布插件，API 成本为 0

I tested DeepSeek Harness for a week. I left with a shipped plugin and $0 in API costs I came to DeepSeek Harness as a reviewer and left as one of its plugin authors, which is a fair measure of how the week went. The hook is on the repo's first line: "Everything is a plugin." Not the usual marketing meaning — not a config file with a couple of third-party hooks. In DSH the agent hosts its own web UI as plugin packages; the settings screen where you pick and configure models is assembled from plugin contributions (this is the part I verified the hard way); even the optional subagents — Codex, Claude Code — ship as plugins you can add or drop. All of it loaded through manifests over a real extension API. I've poked at a lot of "extensible" tools. This one takes it literally, and poking at it turned into the most fun I've had with an agent harness in a while. The test project: a pricing board inside the harness itself The gap I kept feeling in Settings → Models was the two questions I ask before switching providers every time: what does this model actually cost per million tokens, and is a cheaper route to the same model available? DSH tells you which models exist. It doesn't tell you what they cost. dsh-model-pricing closes that, and it lives inside the harness's own Settings page — because that's the point of testing the plugin system: per-1M-token input / output / cache prices for 7,178 priced models across 213 providers (models.dev data) the harness's own routing catalog overlaid for the 788 models it can actually invoke — where the two sources disagree, the row is flagged with ⚠ rather than silently averaged same-model, cross-provider comparison with a cheapest badge: GLM-5.2, say, runs from $3.08 to $4.84 per 1M output tokens depending on the gateway — the same model, 57% apart capability tags (coding / agentic / vision / long context / open weights) computed from structured catalog fields, with a user-configurable rule engine subscription plans (Coding Plan, Token Plan) grouped separately and excluded from per-token ranking a community promotions feed — the part I most want to tell you about, below English + Chinese UI (the platform's) plus a Russian pack the plugin itself registers It installs through the harness's own plugin manager — I never forked anything: npx -y @deepseek-ai/dsh plugin --profile web add dsh-model-pricing Enter fullscreen mode Exit fullscreen mode What surprised me: I built it on a free model Here's the part I didn't expect. I built the whole thing using DeepSeek Harness, driven by Qwen3.8-Flash through B.AI's current promotion — the one the plugin itself ships as a record: "Qwen3.8-Flash API at 0 Credits (free)," with a source link, a verification date and "end not announced" stamped in. Total LLM spend for the project: $0. Which made the week a perfect loop — an agent harness, used to build a plugin about the economics of agent harnesses, on a free model the plugin advertises. Eight records shipped on day one; one of them describes the free model that wrote this plugin's UI. Numbers, since the internet likes them priced rows / providers 7,178 / 213 harness-catalog overlaid rows 788 (30 disagreements flagged, not averaged) plan rows grouped & excluded from ranking 175 promo records at launch 8 (each with source URL + verification date) commits / files / LOC 28 / 18 / ~2,300 wall-clock ~22h, incl. overnight autonomous rounds LLM cost $0 — free API promotion, re-verify date 2026-10-10 UI languages en + zh (platform), ru (plugin) Try it — as reviewer and as builder If you're curious about DeepSeek Harness the way I was: run it, then break into its settings page. The architecture is the product; the agent is very good at both jobs, even on a model that costs nothing. npx -y @deepseek-ai/dsh plugin --profile web add dsh-model-pricing Enter fullscreen mode Exit fullscreen mode Repo (docs, feature map, the promotion contribution rules — all public): github.com/vitas/dsh-model-pricing · npm: dsh-model-pricing · Apache-2.0.