Model providers

The reasoning model runs behind the Scopwis Server. Clients never need a provider API key, and provider credentials are not returned through Bootstrap, Agent observations, reports, or public shares.

Supported transport contracts

Scopwis supports four standard model transports:

• openai-compatible — Chat Completions;
• openai-responses — OpenAI Responses;
• anthropic-messages — Anthropic Messages;
• gemini-generate-content — Gemini GenerateContent.

Model IDs are opaque strings. Scopwis does not compile a model-name allowlist into planning or policy.

The Provider Catalog contains presets for common global, Chinese, cloud, gateway, and local runtimes. A preset is configuration metadata, not an authorization bypass; custom endpoints can use a supported standard protocol even when there is no named preset yet.

Configure from Settings

For a known provider:

1. open Settings → Models;
2. choose the provider;
3. enter the token or credential;
4. load available models when the provider supports discovery;
5. choose a model;
6. run Test;
7. Save.

Protocol, base URL, authentication mode, timeout, and output-token settings remain under advanced configuration.

The test is non-persisting. Save encrypts the configuration and hot-reloads it for new Agent runs. An already-running task keeps the provider it started with.

Environment configuration

Headless deployments can use:

SCOPWIS_MODEL_PROTOCOL=openai-compatible
SCOPWIS_MODEL_AUTH=bearer
SCOPWIS_MODEL_API_KEY=...
SCOPWIS_MODEL_BASE_URL=https://api.example.com/v1
SCOPWIS_MODEL_NAME=provider/model-id

SCOPWIS_MODEL_AUTH accepts bearer, api-key, or none. No-auth mode and plain HTTP are intended only for trusted loopback model servers.

Usage metrics

When Anthropic reports prompt-cache creation or reads, input-token metrics include those counts as well as uncached input. Missing usage is not estimated from text. Malformed known counts or an overflowing input sum fail provider acceptance rather than publishing a partial result. These metrics are not a price estimate or a new run-token limit.

Compatibility requirement

Production Agent execution requires native function/tool calling. A model that can only emit prose cannot satisfy the structured submit_answer and business-tool contracts.

Scopwis does not follow model HTTP redirects, and remote base URLs must not contain credentials, query strings, or fragments.