dsh-standardcompute

A DeepSeek Harness (dsh) plugin that adds Standard Compute as a model provider.

Standard Compute is an OpenAI-compatible API (https://api.stdcmpt.com/v1) with one key for many models. Its standardcompute model is a router: for each request it picks a model and provider, sending routine steps to cheaper models and harder ones to stronger models. You can also pin one exact model by id.

This package is configuration only. Requests go through dsh's own @deepseek-ai/dsh-llm-pi-ai adapter; the bundle adds one route to it. There is no code that runs on your machine besides dsh itself.

What you get

A provider route named standardcompute with three models in the picker:

Model id	What it is
standardcompute	The router. Picks the model per request. Text and images.
deepseek/deepseek-v4-pro	Pinned DeepSeek V4 Pro. Text.
anthropic/claude-sonnet-5.5	Pinned Claude Sonnet 5.5. Text and images.

Any other id from GET https://api.stdcmpt.com/v1/models works too: add it on Settings → Models (edit the Standard Compute provider, Fetch available models) or in your profile's cordis.patch.yml.

Install

Requires dsh 0.2.0-rc.1 or newer and Node.js ^22.19.0 || >=24.

From the release tarball (no npm account or build step needed):

dsh plugin --profile web add https://github.com/henrikdukefoss/dsh-standardcompute/releases/latest/download/dsh-standardcompute.tgz

Or straight from GitHub:

dsh plugin --profile web add github:henrikdukefoss/dsh-standardcompute

In the desktop app or Web UI you can paste either spec into Plugins → Add plugin.

Then restart the profile (dsh web) and give it a key. Create one at standardcompute.com/dashboard, then either:

• open Settings → Models, edit Standard Compute, paste the key and save (stored write-only in $DSH_HOME/.credentials.yaml), or

• export it before starting dsh:

  export STANDARDCOMPUTE_API_KEY=...
  

Pick Standard Compute (router) in the composer's model menu. Selecting it also makes it the default for new sessions.

Check what the bundle added without starting anything:

dsh --profile web --dump-config   # shows "# == @deepseek-ai/dsh-base, patched by dsh-standardcompute"

Using Magpie? One click adds Standard Compute as a provider (you paste your key in Magpie). Guide: https://standardcompute.com/integrations/magpie

How the router works

You send model: "standardcompute". The gateway looks at the request and picks a model and an upstream provider for it; routine steps go to cheaper models and harder steps to stronger ones. A pinned id such as anthropic/claude-sonnet-5.5 skips that choice and always uses that model. Images are accepted on the router; it chooses a vision-capable model when a request carries one.

Reasoning: this bundle declares no reasoning levels, so dsh sends none and the reasoning setting on your Standard Compute dashboard applies. If you want dsh's Effort menu instead, declare the levels on a model (dsh then sends reasoning_effort, which overrides the dashboard setting for that request):

- id: standardcompute
  name: Standard Compute (router)
  contextWindow: 1000000
  input: [text, image]
  reasoningEfforts:
    low: low
    medium: medium
    high: high

Billing: requests stop at the budget

Standard Compute is a monthly plan with a fixed compute budget, not pay-per-token. When the plan's budget is used up, new requests fail with HTTP 402 until the plan renews or you change plans. Requests already running finish, and nothing is billed on top of the plan. In dsh a 402 shows up as a failed request on the standardcompute route; switch to another provider in the model menu if you need to keep going. Plans and budgets: standardcompute.com.

Already have other providers?

A Cordis patch replaces a row's whole config, and your profile's own cordis.patch.yml is applied after plugin bundles. So if that file already configures llm-pi-ai (for example because you added a provider by hand before installing this plugin), your file wins and the standardcompute route does not appear. Two ways to fix it:

• Paste the standardcompute: block from this package's cordis.patch.yml under providers: in your profile's llm-pi-ai entry, or
• add it by hand on Settings → Models → Add model provider → Custom model API: Provider ID standardcompute, base URL https://api.stdcmpt.com/v1, protocol OpenAI Chat Completions, model standardcompute.

Adding providers on the Models page after installing this plugin is fine: dsh's config editor writes the complete current provider list, standardcompute included, into your profile.

Configuration written by this plugin

- id: llm-pi-ai
  config:
    providers:
      standardcompute:
        displayName: Standard Compute
        apiKeyEnv: STANDARDCOMPUTE_API_KEY
        api: openai-completions
        baseURL: https://api.stdcmpt.com/v1
        compat:
          supportsDeveloperRole: false
          maxTokensField: max_tokens
        models: [standardcompute, deepseek/deepseek-v4-pro, anthropic/claude-sonnet-5.5]  # abbreviated

apiKeyEnv is a credential reference, not a key. Nothing secret is stored in this package.

Troubleshooting

• MISSING_CREDENTIAL: no key found. Save it on Settings → Models or export STANDARDCOMPUTE_API_KEY.
• AUTH: 401 Invalid API key: the key is wrong or revoked. Create a new one on the dashboard.
• HTTP 402: the plan's compute budget is used up for this period. See "Billing" above.
• The route is missing after install: see "Already have other providers?".

More setup notes and how this compares with a DeepSeek key or OpenRouter: Best API for DeepSeek Harness.

Development

npm install
npm test                               # manifest and patch checks
DSH_BIN=$(command -v dsh) npm test     # also installs into a temp DSH_HOME and checks --dump-config (needs pnpm)
npm run check:models                   # every bundled model id is still listed by /v1/models (public, no key)
npm run pack:check                     # what would ship

Publishing to npm (npm publish) runs the tests first via prepack.

Disclosure

This plugin is maintained by Standard Compute. It is not affiliated with or endorsed by DeepSeek.

License

MIT