dsh-ollama-cloud

Ollama Cloud for DeepSeek Harness (dsh). One API key; no local Ollama install.

English | 中文

Install

# desktop app profile and/or web profile
dsh plugin --profile desktop add git:github.com/pei-lin-001/dsh-ollama-cloud
dsh plugin --profile web add git:github.com/pei-lin-001/dsh-ollama-cloud

Open Settings → Ollama Cloud, paste a key from ollama.com/settings/keys, and press 保存并同步 (save and sync). The key goes into the credentials service ($DSH_HOME/.credentials.yaml) and never reaches the browser. The catalog is written to the llm-pi-ai settings section as route ollama-cloud, and the models appear in the picker.

OLLAMA_API_KEY in the process environment also works. It shadows a stored key, and the credentials service then refuses writes.

Development checkout:

git clone git@github.com:pei-lin-001/dsh-ollama-cloud.git ~/Dev/dsh-ollama-cloud
cd ~/Dev/dsh-ollama-cloud && dsh plugin --profile web add link:$PWD
./scripts/link-peers.sh   # schemastery, dsh-credentials, dsh-settings, dsh-tools

A linked package resolves peers from its own node_modules, which is why the last step exists.

What it does

On boot, on credential change, and on demand, the plugin pulls the account's catalog and writes it as an llm-pi-ai route: GET /v1/models for the ids, then one POST /api/show per model. Chat, tools, vision and thinking are served by dsh-llm-pi-ai (api: openai-completions, baseURL: https://ollama.com/v1); the plugin supplies the catalog and the credential, and dsh-llm-pi-ai does the transport.

Model entries come from /api/show:

• capabilities decides which models get tools, vision and a thinking set. Models without tools are skipped unless includeNonToolModels is on.
• <arch>.context_length becomes contextWindow.
• thinking.values becomes reasoningEfforts. false maps to off: 'none', since sending nothing would fall back to the model default (usually on). true maps to a single medium level. Graded names map one-to-one, so a model advertising both xhigh and max offers both.

The list replaces the route's models, and an unchanged section is not rewritten.

Two tools come with it: ollama_web_search and ollama_web_fetch call /api/web_search and /api/web_fetch with the same key.

Quota and usage come from GET /api/balance (5h session, 7d weekly) and GET /api/usage. A current plan reports tokens and cost there; a legacy plan reports requests only. For legacy plans the plugin taps the usage chunk its own requests emit through llm/stream, counts input/cached/output per model in hourly buckets, keeps 30 days in $DSH_HOME/ollama-cloud-usage.json, and aligns the totals with the reset times the balance reports. Other clients' traffic stays invisible to that meter, so server numbers win when they exist.

Every history image is re-sent on every step, and Ollama Cloud rejects an oversized request at the gateway before the model runs (413, around 12 MB observed). The plugin keeps image base64 under 6 MiB: on 0.2.x it writes maxRequestImageBytes into the route, so the adapter stops at the bound and the harness offloads the oldest images and retries with placeholders; on 0.1.0-rc.x, whose adapter has no such bound, the plugin drops the oldest images from the outgoing request itself. Text is not part of that budget.

UI

The settings page has four parts: connection, quota bars, charts, per-model table. The 24h chart is hourly requests. The 7d chart switches between requests and tokens, because requests are the series every plan returns. The sidebar footer shows both quota windows. UI copy is Chinese.

Configuration

Entry config is the ollama-cloud settings section; override it in the profile patch:

- id: ollama-cloud
  config:
    provider: ollama-cloud        # llm-pi-ai route name
    displayName: Ollama Cloud
    apiKeyRef: OLLAMA_API_KEY     # credential reference, never the secret
    baseURL: https://ollama.com
    syncOnStart: true             # pull the catalog on boot
    includeNonToolModels: false   # keep tool-less models out of the picker
    webTools: true                # register ollama_web_search / ollama_web_fetch
    maxOutputChars: 20000         # per ollama_web_fetch result
    usageCacheSeconds: 120        # /api/balance + /api/usage cache
    maxRequestImageBytes: 6291456 # base64 image budget per request; 0 = adapter default

A model's contextWindow comes from /api/show. The composer's 上下文 chip beside the model button overrides it for the selected model only; the choice is kept in $DSH_HOME/ollama-cloud-context.json and layered on top of every catalog sync. Resetting restores the reported value.

Command and routes

/ollama-cloud [status|sync|usage], defaulting to usage:

Ollama Cloud 用量：
  会话（5h） ▕█░░░░░░░░░▏ 已用 7.2% · 3 小时后重置
  每周（7d） ▕█░░░░░░░░░▏ 已用 6.4% · 18 小时后重置
  5h 统计：267 次请求 · 输入 8.9k · 输出 127
  7d 统计：1.2k 次请求 · 输入 8.9k · 输出 127

The settings page and the sidebar use these routes over the same origin; scripts can too.

Method	Path	Purpose
GET	/_dsh/ollama-cloud/status	key configured?, route, model count, last sync
GET	/_dsh/ollama-cloud/usage[?force=1]	balance windows, usage, token totals
POST	/_dsh/ollama-cloud/key	{ "key": "…" } stores and syncs, { "remove": true } deletes
POST	/_dsh/ollama-cloud/sync	re-pull the catalog

/key and /sync write. They take the app's own page only (same origin, or an opaque file:// origin) and a loopback or IP-literal Host; everything else gets 403.

Runtime compatibility

Runtime	Settings seam	Behavior
0.1.0-rc.x	ctx.settings.register / get	the plugin owns the ollama-cloud section; the adapter's section is read with get('llm-pi-ai')
0.2.x (desktop)	describe / mutate	the entry config is the section; the adapter's entry is read with describe() and written with mutate()

engines.dsh accepts >=0.1.0-rc.6 <1.0.0, so both runtimes load the bundle. Tested on 0.1.0-rc.6 (CLI) and 0.2.1-rc.2 (the desktop app's bundled runtime). CSS tokens carry literal fallbacks, so the older runtime renders the same layout with plainer colors.

Known limits

There is no way to cap output from here. pi-ai's compatibility probe sends max_completion_tokens, which Ollama ignores (it honors max_tokens), and dsh-llm-pi-ai exposes no switch for that field.

The local meter only sees traffic from this harness home. Another machine or another client will not appear in the token totals, and hourly buckets can put a window edge off by up to an hour.

When the catalog changes, the sync replaces llm-pi-ai.providers.ollama-cloud whole, so any hand edits to that route are lost. syncOnStart: false makes syncing a manual step.

webTools and includeNonToolModels are read when the tools register, so changing them needs a restart. Client-side changes also need one.

The usage endpoint allows 10 requests per minute per user. One refresh is three calls, which is why the result is cached and the sidebar polls every five minutes.

Development

node test/smoke.mjs   # catalog mapping, thinking levels, usage math, meter, route guards

No build step. The host half is plain ESM; the client half is written in the loader's factory form (window.__ModuleLoader__.load({ id, factory })).

To run the web profile on the desktop app's bundled runtime:

APP="/Applications/DeepSeek Harness.app/Contents/MacOS/DeepSeek Harness"
CLI="$APP/../Resources/app.asar/dsh/node_modules/@deepseek-ai/dsh-desktop-host/lib/cli.js"
ELECTRON_RUN_AS_NODE=1 DSH_HOME=/tmp/probe "$APP" --expose-internals "$CLI" --profile web --port 8130 --no-open

License: MIT.