Practical fixes for the things that actually break: usage limits mid-task, 401s after you switch endpoints, model-not-found errors, and agents that silently keep using the wrong model. No signup, no email gate.
Each of these is a specific failure with a specific cause. Guessing wastes an afternoon.
Answer three questions, get the specific cause and the exact commands. Runs on rules, not a chatbot — same answer every time.
errorYour key is probably fine. Usually the process never saw the variable, or the gateway wants a different auth header.
errorAuth succeeded, resolution failed. Model slugs differ between vendors and routers. Stop guessing the name — list it.
capacityCut consumption first — it's free. Then split work by difficulty. Then, if you still need it, configure a fallback endpoint.
configBase URL and model are independent settings. Setting one and not the other is the single most common config mistake.
decisionSubscription vs API vs router. Measure your real token volume before you buy — most people over-buy.
Most "fix your AI coding setup" content is a wrapper around an affiliate link. The advice is vague because vague advice converts about as well as good advice, and it's cheaper to write.
We took the opposite approach: the diagnostics are deterministic and the commands are real, because a page that actually fixes your problem is the only kind worth coming back to. We do earn referral rewards when someone signs up to AgentRouter through us — that's stated on every page and it's why this is free. What we won't do is quote a price or a promo we haven't checked, invent a benchmark, or tell you to switch providers when your actual problem is a missing environment variable.
Before anything else, find out which of the three layers is broken. Whichever fails first is your real problem.
| Layer | Test | What a failure means |
|---|---|---|
| Reachability | curl -sv <base-url>/v1/models |
Connection refused or timeout — wrong host/port, nothing listening, or a doubled /v1 path. |
| Authentication | Same call with your auth header | 401/403 — wrong credential, wrong header style, or whitespace captured in the variable. |
| Model resolution | A real completion request | 404 / model not found — the slug isn't valid on this endpoint. List them and copy one verbatim. |
If curl passes all three but your tool still fails, the tool isn't seeing your environment.
That's a shell inheritance problem, not a provider problem — restart the editor from a shell
where env | grep -i api shows what you expect.
Most problems on this site are configuration, and the pages above fix those for free. But if you have genuinely run out of capacity, or you need a model your current endpoint doesn't serve, you need somewhere to point your tools. AgentRouter is OpenAI-compatible, publishes per-token pricing you can check before committing, documents setup for 15 tools, and serves gpt-5.6-sol, claude-opus-4-8 and claude-opus-5 on its default group.
Look at the setup guides →Referral link — we may earn a reward if you sign up, at no extra cost to you. This is the same advice we would give with no link at all. Their own docs note the service targets developers rather than guaranteed uptime, so check current pricing, model availability and status on their site before you depend on it.