Model providers
English | 中文
Autoloom does not include model usage. You choose a provider, and that provider controls model access, billing, rate limits, data handling, and availability.
Add a provider
Open Settings → Models.
- Use a built-in catalog provider when your provider is listed. Enter the credential that provider requires and save.
- Use Accounts for a provider that exposes a supported sign-in or subscription flow. Opaque grants are stored in Windows Credential Manager.
- Use Add a custom provider for a company gateway, self-hosted service, or unlisted OpenAI-compatible endpoint. Supply a stable lowercase provider ID, base URL, protocol, credential, and at least one model.
A custom provider ID becomes part of saved model selections and credential references. Add a new provider instead of renaming an existing ID.
API keys saved in the Models page are write-only in the UI: the Renderer receives a redacted descriptor, not the literal value. They are stored in the managed credentials file under the Autoloom data directory. See Data and privacy.
Select a model
Choose a default model after saving the provider. A Session can select another model explicitly; Follow global returns it to the current global default. A changed global default applies to the next request of Sessions that follow it and does not change a request already in progress.
If a saved Session points to a provider that has been removed, select another model before sending a new message.
Custom endpoints and images
An endpoint can accept the right key and still reject a request because its API differs from the selected protocol. Confirm the provider's documented base URL and protocol. If a gateway only partially implements an OpenAI-compatible API, its request-role, token-limit, or reasoning fields may need provider-specific compatibility settings that the current form does not expose.
A model entered by hand is treated as text-only unless its configuration declares image input. Autoloom refuses an image before dispatch when the selected model does not declare image support. Declaring support tells Autoloom what to send; it does not prove the endpoint can process the image.
What is sent
For each turn, Autoloom sends the assembled request to the selected endpoint. It can include system instructions, conversation history, tool definitions and results, selected file content, and attachments. Credentials authorize the request but are not inserted into the conversation. Review the provider's own privacy and retention policy before using sensitive projects.
Web search and page reading
By default, web_search uses Parallel's public MCP endpoint without a separate search account or API key. Exa's public MCP endpoint is the alternative; a transient failure or rate limit permits one attempt with the other public service. Anonymous access has upstream quotas and availability limits. Autoloom does not silently switch to a paid service. Only the search query is sent to these services, not your conversation, project files or model credentials.
Local and SSH projects both support default search. SSH search and page reading run on the remote host and depend on that host's network access. Search does not require the chosen model to support hosted search. The optional model mode uses verified hosted search on the selected route and an explicitly configured independent fallback.
web_fetch reads a specified HTTP(S) URL directly. If a proxy's Fake-IP DNS returns a synthetic address for a hostname, Autoloom obtains real addresses through HTTPS DNS and validates that they are public before connecting. Literal reserved addresses, private destinations and unsafe redirects remain blocked. You do not need to change your proxy's DNS configuration for this path.
Search results and fetched pages are supplied to the Agent as untrusted reference data, so instructions embedded in a page do not override your request or Autoloom's rules. Your chosen model still consumes tokens when it processes these results.
Common failures
- Missing credential: save the required key or sign in to the required account, then retry.
- Unknown model: select a listed model or add the exact model ID to the custom provider.
- 401 or 403: check the credential, account permissions, endpoint, and provider region or project settings.
- 429 or quota error: review the provider's rate limit, balance, and billing status.
- Connection or timeout error: confirm the endpoint is reachable from the Windows computer; SSH projects still dispatch model requests locally.
- Every request is rejected by a custom gateway: recheck the chosen protocol and the gateway's compatibility with system/developer roles, reasoning fields, and token-limit fields.
- Image rejected before sending: choose a catalog model with image support or configure the custom model's input capability.
Never post a live key or unredacted provider error payload in a public Issue.
