dsh-llm-subscription
适用于 Claude Code / Antigravity CLI 订阅的原生 DeepSeek Harness LLM 适配器，支持本地 Ollama——在模型选择器中使用 Claude、Gemini 和免费的本地模型，无需 API 密钥
插件会安装到这里；不确定时保持 web。
npx -y @deepseek-ai/dsh plugin --profile web add dsh-llm-subscription@0.1.4Llm Subscription 以 dsh-llm-subscription 发布，当前版本为 0.1.4。Plugin Hub 会校验它的 manifest，并保存精确安装来源，便于复现安装结果。
正在加载相关插件…
A native DeepSeek Harness LLM adapter that puts Claude and Gemini in the model picker — real entries, right next to DeepSeek, with a working reasoning-effort selector — using your existing Claude Code / Antigravity subscriptions instead of a metered API key. It also registers a local Ollama model, for a $0 option with no login of any kind.
If you've seen dsh-subscription-gateway, this is the same idea done properly: that one is an HTTP bridge you point DSH's Custom Provider at. This one is an actual dsh.bundle plugin registered against DSH's own dsh-llm seam (ctx.llm.registerAdapter) — same mechanism dsh-llm-deepseek uses internally. No Custom Provider screen, no separate process to keep running; it's just there once installed.
claude, haiku, sonnet, opus, fable — each with a native reasoning-effort selector (Low → Max), because Claude's own --effort flag supports that independently of which model you pick.Flash Low/Medium/High, Pro Low/High) — Gemini bakes its intensity into the model name itself, so there's no separate effort control for it.qwen3.5:9b by default) — no CLI login, no subscription, nothing to authenticate. Starts ollama serve itself the first time it's actually used if it isn't already running; never touches it otherwise.via <real-model-id> line — claude --model sonnet is an alias, and Claude Code sometimes uses a small internal model alongside it for things like session titling. This picks out whichever model actually wrote the visible text (the one with the most output tokens) so you're not left guessing whether you got Sonnet or something else.Inside a DSH profile:
dsh plugin --profile web add dsh-llm-subscription
Requires at least one of: the claude CLI (logged in via claude login), agy (Antigravity CLI, logged in), or Ollama with a model pulled (ollama pull qwen3.5:9b) — all already on PATH. Missing one just means that provider's models won't show up; the rest still work fine.
Claude and Gemini requests shell out to claude -p or agy -p in headless/JSON mode and reshape the result into the StreamChunk sequence DSH's LLM seam expects (block-start → text-delta → block-end → usage → finish). Neither CLI streams token-by-token in headless mode, so the whole response arrives as one delta — functionally fine, just not incremental. Ollama requests go straight to its local HTTP API (/api/chat) the same way, also as one non-streamed delta, for consistency with the other two.
The Claude call always passes --strict-mcp-config. Without it, the subprocess quietly inherits whatever MCP servers are configured in your regular Claude Code setup, which leaked unrelated tools into responses that were supposed to be clean completions — a real bug hit while building this, not a guess.
It might — cordis and the dsh-llm seam are developer preview with no stable contract, and this plugin depends on internals discovered by reading dsh-llm-deepseek's compiled source, not from documentation. dsh-subscription-gateway does the same job through DSH's stable, documented Custom Provider screen instead — worse UX (no native picker, no effort selector), but it doesn't touch anything that changes between DSH releases. Worth having as a fallback.
agy's own built-in tools the way Claude calls are (--strict-mcp-config) — there's no equivalent flag. Verified empirically: asked headless, agy reports its full native tool set. This adapter never parses or acts on tool-call output, so there's no execution path through here, but the model's own awareness of those tools can still shape its answers.0.1.1-rc.2. This whole plugin surface (cordis, the dsh-llm seam) is developer preview and can change without warning.This never reads, stores, extracts, or transmits any credential. It shells out to the claude / agy CLI binaries already installed and logged in on your machine, and reads their stdout. Nothing is shared, proxied, or routed between users — each install runs against your own already-authenticated session.
In February 2026, Anthropic explicitly banned third-party tools (OpenClaw, NanoClaw) that extracted a Claude subscription's OAuth token and reused it to authenticate a separate, direct API client — bypassing Claude Code entirely. That's not what this does. Anthropic's own guidance: "OAuth authentication is intended exclusively for purchasers of Claude Free, Pro, Max, Team, and Enterprise subscription plans and is designed to support ordinary use of Claude Code and other native Anthropic applications." What's explicitly prohibited is reselling or intermediating Claude usage between users — each end user authenticating with their own credential is the compliant pattern, and that's what happens here.
Update (checked August 2026): Anthropic's own April 2026 clarification of that ban states plainly that "running the official CLI remotely... is explicitly supported" — this plugin shells out to the real claude/agy binaries, the same pattern, not the OAuth-token-reuse pattern that got banned. Worth knowing this is an area Anthropic is actively watching, not settled forever: in June 2026 they announced moving claude -p usage specifically to a separate, smaller monthly credit pool — then paused it before it took effect, so claude -p still draws from your normal subscription today. Not legal advice, no guarantee this stays true. If you're installing this for anything beyond personal use, read Anthropic's Usage Policy yourself and keep an eye on their own announcements.
Standard this project follows for any provider it adds: never read/cache/transmit a credential on the user's behalf; only ever invoke the vendor's own official CLI in a documented automation mode; never implement a login flow ourselves; no usage pooled or shared across users.
MIT — see LICENSE.