FAQ
Provider setup and troubleshooting for cloud and local agents.
Provider setup & troubleshooting
AICoven is BYOK: you bring provider credentials or a local agent endpoint. A provider can be saved and reachable yet still fail a real chat turn if the model, deployment, or connection is wrong. These are the most common causes.
I added a provider key — why do I still get no reply?
Saving a key proves only part of the setup. A real chat turn also needs the role you're chatting with to be assigned to that provider account, a chat-capable model or deployment, and billing/credits on the provider side. If a turn fails, open the error card on the message — it says whether the key is invalid, credits are insufficient, the model isn't a chat model, or no working chat provider was available for the turn.
Why is Azure OpenAI set up differently from OpenAI?
Azure OpenAI uses your Azure resource endpoint and named deployments rather than public model names. Where AICoven asks for the model/deployment, enter your Azure deployment name (e.g. prod-gpt-4o-chat), not just the model family like gpt-4o. If AICoven says the provider connected but returned an empty chat response, the deployment name or its chat capability is usually the culprit.
Why does localhost work in my browser but not in AICoven?
AICoven's API runs in the cloud, so localhost refers to the cloud container, not your machine. For private endpoints (localhost, LAN, .local, or Tailscale addresses), keep the AICoven Mac app connected — it relays requests from the cloud to your machine. Alternatively, expose an authenticated HTTPS endpoint. Read more in how we route cloud APIs, Hermes, OpenClaw, and your Mac.
Do I have to type my local server's address?
On macOS, usually not. Add Provider Key has a Detect button that checks the ports local servers normally use — OpenClaw on 3333, LM Studio on 1234, the Hermes local relay on 8642, Ollama-compatible servers on 11434 — and anything it finds connects with no typed fields.
Moved OpenClaw or the Hermes relay off its default port? Grant AICoven access to its config folder and Detect reads the real port — and the auth token — straight from that config. LM Studio and Ollama don't record their port anywhere AICoven can read, so enter it yourself in the “Running on another port?” field, name the server, and pick how to route it. AICoven never scans ports: it reads a port where the tool records one, and asks you where it doesn't.
Can AICoven use the Google Cloud or AWS credentials I already have?
On macOS, yes — for Vertex AI and Amazon Bedrock. If you've run gcloud auth application-default login, Detect reads those credentials and picks up your project from the gcloud config. If you've configured the AWS CLI, it reads a profile from ~/.aws along with its region, and lets you choose which profile when you have several.
Both ask you to grant folder access once, through the standard macOS folder picker, and AICoven reads only that one credentials file. Two limits worth knowing: AWS profiles that use IAM Identity Center (SSO) keep no keys on disk, so they can't be detected and still need entering another way; and on iPhone or iPad there's no CLI to read from, so paste the credentials instead.
Which URL should I use for Hermes or OpenClaw?
Point AICoven at the agent runtime itself — the Hermes Agent gateway or your OpenClaw service — not at a generic OpenAI-compatible chat server. Agent roles use the durable run lifecycle (runs, approvals, continuation), which a bare chat-completions endpoint doesn't support.
Hermes asked for approval — why doesn't typing "approved" work?
Approvals belong to the running agent run, not the chat. Use the approval controls on the approval card when it appears; a typed chat message starts a new turn and won't unblock the pending action.
Why can't Hermes or OpenClaw see my uploaded file?
Uploads live in AICoven's app flow, not automatically on your agent's host machine. AICoven relays extracted text or document content into the prompt where possible. Prefer text, Markdown, or PDF/DOCX with extractable text — and if the agent should operate on a file on your Mac, make sure the local agent connection is active and the file exists on that machine.
Chat works — why do image or voice tools fail?
Chat, image generation, speech transcription, web search, and local tools are separate capabilities. Add a provider account that explicitly supports the capability you need, and select a model for it where the UI asks for one.