Bring your own provider
SPM does not resell model inference. You keep an upstream provider account and choose one of two credential-custody models.
Credential custody
| Hosted Provider Proxy | Local Proxy | |
|---|---|---|
| Provider key storage | Encrypted SPM vault | Protected local config |
| Model request path | Client → SPM → provider | Client → loopback Local Proxy → provider |
| SPM memory path | Hosted | Hosted |
| Hosted gateway receipts | Yes | No |
| Model selection | Request model routed against allowed models | Downstream harness sends the model |
Neither path is universally safer. Hosted forwarding centralizes operations and receipts. Local Proxy reduces provider-key disclosure to SPM but adds local configuration and process-security responsibilities.
Native API styles
| API style | Endpoint |
|---|---|
| OpenAI Chat Completions | POST /v1/chat/completions |
| OpenAI Responses | POST /v1/responses |
| Anthropic Messages | POST /v1/messages |
SPM is not a cross-dialect translator. Requests and provider responses stay in the configured native style.
Hosted provider configurations
One active provider configuration represents one upstream Base URL, credential, API style, custom-header/query set, default model, and allowed_models[].
- A request with an explicit model routes to the configuration that owns that model.
- A request without a model uses the default channel/model for that API style.
- Two active configurations under the same tenant and API style cannot claim the same model.
- One upstream configuration can serve multiple models; do not create a new provider row solely to change model.
Presets, auto-admission, and models.dev
The provider catalog is backed by models.dev and refreshes automatically with stale-while-revalidate plus a committed snapshot fallback.
Two tiers of providers appear in Settings:
- Preset providers (major providers such as OpenAI, Anthropic, DeepSeek) carry curated metadata — console links, docs, and connection details.
- Auto-admitted providers: when the live catalog lists a provider with a working HTTPS endpoint, a documented credential variable, and at least one chat-capable model, it becomes selectable without waiting for a review cycle. Pricing (input/output per 1M tokens), context window, and capability markers (reasoning, tool calling) refresh with the catalog.
Auto-admission is not a safety bypass: every request still goes through the same HTTPS-only, public-IP-pinned, SSRF-checked transport, and an auto-admitted channel stores its catalog-provided endpoint at creation time. Providers whose catalog entry is incomplete (no usable endpoint or credential variable) remain discovery-only and are never routable.
Preset providers do not live-fetch /models. This keeps the catalog stable and avoids provider-specific permission/cost surprises.
Custom BYO
Use Custom for a self-hosted gateway, regional provider, enterprise proxy, or provider not covered by a preset.
You configure:
- HTTPS Base URL
openai_compatibleoranthropic_messagesAPI style- provider key
- default model and allowed models
- optional custom headers and query parameters
Only Custom channels can use Fetch models. For an existing channel, the server-only BFF reads the credential temporarily from the vault, performs the guarded probe, and returns model identifiers—not the credential—to the browser. If the upstream has no model-list endpoint at all, you can also type a model ID in manually.
Custom probes and forwarding enforce HTTPS and public-address DNS pinning. SPM, cookie, forwarding, Cloudflare/client-IP, and hop-by-hop headers are removed after caller and custom headers are merged; provider authentication is injected last.
Protocol transparency
Unknown provider JSON fields and valid diagnostic headers pass through by default. Responses SSE event order and response/item/part IDs are not rebuilt. Anthropic thinking and OpenAI reasoning state stay in provider traffic.
The separate memory observer captures only visible assistant text. Reasoning, thinking, redacted thinking, function/tool arguments, and partial JSON are not written as memory.
Revocation and limits
Revoking a key or provider configuration removes it from the console and stops it from authenticating or routing. Provider-channel limits are enforced by plan; deleting a provider frees its slot.
Provider retention and training
The upstream provider's own retention, training opt-out, regional-processing, and account settings apply to model traffic. SPM cannot change those settings for you. Review the provider's terms before sending sensitive content.