loongport

loongport is a DeepSeek Harness (DSH) Cordis bundle for verified LoongPort OpenAI-compatible providers. It also includes an advanced CLI for custom endpoints.

Install in DeepSeek Harness

Install the public package into the DSH profile you use:

dsh plugin --profile <profile> add loongport

Open Settings → LoongPort in DSH, then:

1. Select a verified provider. The default is DeepSeek's official API at https://api.deepseek.com/v1; BestAPI (https://api.bestapi.store/v1) is the verified relay option. 2. If you need an account, use the provider's Register or Sign in link. It opens the provider's own website in your browser. 3. Generate an API key with the provider and paste it into the dialog. 4. Select deepseek-v4-flash or deepseek-v4-pro, then save.

The signed LoongPort directory is the authority for provider identity, URLs, available models, and any published invitation code. VeriDrop data, when shown, is only a display-only health observation and cannot change provider configuration. DSH credentials store API keys; the Settings page only retains whether a key is configured.

Version 0.2.0 does not automate browser authorization, collect browser sessions, or access cookies/local storage. Registration/login and API-key creation remain on the provider's site.

Advanced custom endpoint

For a manually supplied OpenAI-compatible endpoint, run the CLI directly:

LOONGPORT_API_KEY='your-api-key' npx loongport dsh setup \
  --base-url https://relay.example.com/v1 \
  --model model-id \
  --write

The API key is read only from LOONGPORT_API_KEY; never place it in a command argument. You can repeat --model to register more than one model.

By default, the command performs a dry run. Add --write to update DSH's configuration. It uses $DSH_HOME when set, otherwise ~/.dsh, and writes:

• settings.yaml: one selected llm-pi-ai.providers.<route> entry
• .credentials.yaml: the selected credential name and API key (owner-only)

Other DSH settings, provider routes, and credentials are preserved. The command uses DSH's built-in llm-pi-ai transport; it does not implement its own streaming, retries, or tool calls.

Options

--base-url <url>              OpenAI-compatible HTTP(S) endpoint (required)
--model <id>                  Model to expose; repeat for multiple models (required)
--route <name>                DSH provider route (default: loongport)
--credential-name <name>      DSH credential key (default: LOONGPORT_API_KEY)
--write                       Apply changes; omitted means dry run

The CLI remains an explicit advanced path. It supports manually provisioned OpenAI-compatible endpoints, but does not automate third-party browser login, relay discovery, desktop-app bridging, or automatic tier switching.

Maintainers

Agents and maintainers should read the [agent guide](docs/maintainers/agent-guide.md) before changing the package and follow the [release runbook](docs/maintainers/releasing.md) before publishing.

License

[MIT](LICENSE)