Models & providers
Connect inference, choose a model and verify a real response.
A provider is the service hosting an AI model; a model is the particular
identifier you select. Inference means asking that model to produce an answer.
For example, connect an Ollama service, then choose one of its installed models.
Provider setup happens in the interface rather than through routine .env edits.
Connect a provider
The instance owner opens Settings → Providers.
- Choose Local · Ollama, Hosted provider, or the advanced compatible-endpoint option.
- For hosted inference, search the catalog, choose a provider marked compatible and review its destination. Enter its token when required. For Ollama, enter its server address.
- Select Save connection. Saving stores configuration; it does not start inference.
- Select Check connection on the saved card. Choose an installed or served model. If listing is unavailable, an entered/catalog ID remains labelled as not live-verified.
- Select Set instance default to use that connection/model for new requests.
A catalog entry is not an implemented adapter
The searchable catalog covers more services than Wrenly currently supports. The working transport is OpenAI-compatible, including configured Ollama. Other native protocols require adapters. A catalog price or model ID does not prove account availability.
Verify inference
Open a project and ask a simple question. A successful model-list check establishes reachability; an actual response establishes inference. Building additionally needs the framework's Docker toolchain.
If the check fails, confirm the saved destination and service availability. If a model is listed but cannot answer, check account access, quota and the exact identifier. Wrenly does not silently switch providers to hide a failure.
Choose a model on this browser
Open Settings → Models. Search by model/provider or filter Local and Cloud, then select a model for new private-project prompts. Use instance default clears the browser override. This does not change other members' preferences or admitted work.
The choice includes both provider and model: two providers can use the same model name. If an older browser preference has no provider, select the pair again; Wrenly will not guess a provider or select the first available model.
If the browser cannot read or save a preference, Wrenly shows a storage error. Allow site storage and choose the model again. A failed save keeps the previous selection; a failed reset does not claim that the instance default was restored.
A missing saved choice remains visible for explicit correction. Refresh discovery checks availability again; older inventory is labelled when refresh fails.
Shared projects use the owner's saved Project model, configured in project settings. A missing or disabled project connection blocks inference instead of falling back to the instance default. Learn about project access.
Set up local inference
Discovery lists models already installed in the configured Ollama service. An unreachable service and a reachable service with no models have different guidance. Install/run Ollama and a model on the intended inference machine, then check that service from Wrenly. The API machine may differ from the browser machine.
Wrenly does not scan unrelated hosts or download large models automatically. The catalog includes an attributed local snapshot; normal browsing does not require internet. The owner can explicitly refresh metadata without retargeting saved credentials.
Manage a saved connection
Replace its token or disable/re-enable it from Providers. Empty token edits retain the saved token. Destination changes require a new connection so queued work cannot be retargeted. Tokens clear from successful form saves and are not returned to the browser. Members use enabled connections without viewing credentials; hosted usage consumes the configured provider account's quota.
Admitted work retains its connection version/model. Token rotation supplies a replacement credential for that same connection. Disabled connections stop execution rather than selecting a replacement automatically.
Import existing server configuration
Under Existing server configuration, review the destination and choose Import
connection. Wrenly copies the credential server-side to the encrypted store without
changing .env or exposing the token. Finish old requests first, check the saved connection,
then select a default and clear outdated browser overrides.
Repeated imports do not rotate tokens or re-enable disabled connections. Imported sources cannot silently switch to another provider.
Back up credentials
Saved provider tokens are encrypted in PostgreSQL. Their separate encryption key is
created in the API's private persistent data directory. WRENLY_DATA_DIR defaults to
.local/wrenly relative to the API working directory; container installations must persist
that location. Back it up with PostgreSQL. Changing the auth signing secret does not
rotate this key.
If the key is missing while saved tokens exist, restore the matching backup. Wrenly refuses to generate a replacement that would make the old tokens unreadable. Token-free local connections do not need a provider encryption key.