Agnai API setup: base URL, key, and the format field

Quick answer: the fields Agnai's Third Party service needs

Agnai's Third Party connection lives inside a preset, not as one global setting. Open Settings, go to Presets (some builds label this AI Settings), create a new preset or edit an existing one, and set Service to Third Party. Four fields decide whether the connection works: Base URL, an OpenAI-compatible endpoint that should stop at /v1, for example https://api.apisrouter.com/v1; API Key, your provider's secret key; Format, which tells Agnai how to shape the outgoing request, set to OpenAI so it builds a Chat Completions payload instead of a raw text prompt; and Model, the exact ID your key can call, such as deepseek-v4-flash. Saving the preset is not the last step. Agnai applies presets per chat or per character, so an older Kobold or native-service preset left active on a chat keeps routing there even after you build a working Third Party preset elsewhere. Assign the new preset explicitly, either as your default or on the specific chat, before assuming the connection is broken.

What each Third Party field is actually doing

Two things ride along on every message regardless of whether those four fields are correct: the character definition and persona you built, and any memory book entries currently active for the scene. Agnai resends both, plus recent chat history, on every turn, the same way JanitorAI resends a character card and SillyTavern resends its prompt template. None of that lives in the Third Party fields, but it is what actually drives the size, and therefore the cost, of every request those fields make possible.

• Base URL: the address Agnai sends every request to. It should stop at /v1 for most OpenAI-compatible providers; Agnai appends the completions path itself, so a URL that already ends in /chat/completions produces a doubled, broken path.
• API key: sent as an Authorization header to the base URL only. Depending on your Agnai version it sits directly in the preset or under a separate registered-keys section in Settings; either way it never appears in the chat itself.
• Format: controls the shape of the outgoing request and how Agnai parses the reply. OpenAI-compatible providers expect the Chat Completions shape, a messages array with roles. Picking a Format built for a different API's response schema is a common cause of a connection that accepts the key but returns garbled or empty text.
• Model: sent as the model parameter on every request. It has to match a real ID from your provider's catalog, not a display name; a mistyped or outdated ID usually surfaces as a model-not-found style error rather than silence.

Which model to point Agnai at, and why

The base URL, key, and Format stay identical across every model you try; only the Model field changes, so the practical move is one saved preset per model you actually use, the same pattern JanitorAI and SillyTavern users settle into. Discussion around Agnai's Third Party option leans toward the same handful of budget models favored on other companion frontends, for the same reason: steady character-card adherence at a low per-token rate matters more in a long-running chat than a benchmark score. DeepSeek V4 Flash is the common starting point for value, with Pro a step up for chats that lean on a big memory book or a group scene. GLM-5 and Kimi K2.6 are solid budget picks with a different prose feel worth rotating in when a chat starts to feel samey. Grok is notable because xAI's published policy explicitly allows mature themes in fiction, which some users weigh for darker storylines. Claude writes the strongest prose on this list, but Anthropic's usage policy prohibits explicit content, so treat it as a SFW option for character voice and description rather than anything else.

What a chat session actually costs

Agnai resends the character definition, persona, active memory book entries, and recent history with every message, so input tokens dominate the bill, the same shape as JanitorAI and SillyTavern. A modest setup with a couple of active memory book entries commonly runs around 4,000 input tokens and 300 output tokens per message. The table below prices that assumption across 1,000 messages, roughly a heavy month of daily chatting. A bigger memory book or a longer retained history pushes the input side up; a leaner one pulls it down. These are APIsRouter catalog rates, from an OpenAI-compatible gateway with pay-as-you-go billing and no subscription: global models are priced 20% below official list, Chinese models sit under their official rates, the first top-up adds a +100% balance bonus, and the /topup checkout skips the signup form, taking payment first and emailing the key.

Why the connection breaks, or the bill jumps, without warning

Test the four fields in isolation before assuming any one of them is the culprit. The curl example later on this page is the fastest way to separate a bad URL or key from a Format problem inside Agnai itself.

• Base URL ending in /chat/completions instead of /v1: Agnai appends the path itself, so the request lands on a doubled URL and fails before it reaches your provider.
• Format mismatched to the endpoint: the connection accepts the key but sends or parses the wrong shape, which reads as a garbled or empty reply rather than an obvious error.
• A mistyped or renamed Model ID: usually surfaces as a model-not-found response, not silence, so it is worth checking first when nothing comes back at all.
• A working preset saved but never assigned: the chat keeps using whichever service was active before, so the new Third Party preset looks broken when it was simply never applied.
• Memory book entries left active from a finished arc: they keep riding on every message, quietly inflating input tokens on scenes that no longer need them.
• Free, queue-based community compute mistaken for a broken key: a full queue on shared backends looks identical to an authentication failure from the chat window, and the fix is unrelated to your Third Party fields.

Third Party key versus Agnai's other connection options

Agnai reaches models three broad ways, and Third Party is only one of them. Shared community compute needs no key at all, but capacity is queued and shared, so response time and availability move with demand rather than with anything you control. A native integration talks to one specific provider directly and is as reliable as that provider, but it locks you to whatever models that integration supports. Third Party is the general-purpose option: any OpenAI-compatible endpoint works once you have a key, at the cost of the one-time setup covered above.

Fixing it: verify outside Agnai, then match the fields

The curl call below is exactly what Agnai's Third Party connection does under the hood: same URL, same header, same JSON shape. If it returns a completion, your Base URL, key, and Model are all valid, and a broken connection inside Agnai points at the Format field or an unassigned preset instead. The second call lists every model ID your key can use, which settles any doubt about spelling before you touch the Model field again.

• Test the endpoint directly with curl before touching Agnai again, so you know whether the URL, key, or model ID is at fault before blaming the app.
• Set Format to OpenAI whenever the provider is OpenAI-compatible. A Format built for a different API is the most common cause of a connection that accepts the key but returns junk.
• Paste the model ID exactly as the provider lists it, lowercase and unabbreviated, and recheck it against the provider's model list rather than a display name.
• Assign the working preset to the chat, or set it as your default, instead of leaving an older service active on that character.
• Prune memory book entries you are not using in the current scene; every active entry rides on every message whether the scene needs it or not.
• Point the same Third Party key at a lower-priced OpenAI-compatible endpoint for everyday chatting, and save a pricier flagship model for the scenes you actually care about.

curl https://api.apisrouter.com/v1/chat/completions \
  -H "Authorization: Bearer sk-YOUR_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "deepseek-v4-flash",
    "messages": [{"role": "user", "content": "Say ready."}],
    "max_tokens": 20
  }'

FAQ