Build with Keplar.

Every answer through the API or a widget is metered exactly like an answer in the app: same credits, rolling windows and spend caps. The per-key and per-widget limits below only ever make things stricter.

1. Create a key

   Sign in, open Account → Developer API, name the key and pick an expiry. It is shown once; we keep only a hash.

2. Keep it on a server

   Store it as KEPLAR_API_KEY. Never ship it in a web page or a repository. For a public page, use a widget instead.

3. Ask

   Send Authorization: Bearer kpl_… with a JSON body. Requests are stateless, so pass earlier turns in history.

curl https://keplar.one/api/v1/ask \
  -H "Authorization: Bearer $KEPLAR_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"question": "Should a two-person startup rewrite its Node.js backend in Rust?"}'

Scopes: ask Ask questions: POST /api/v1/ask (including streaming). Spends your plan's credits. usage Read usage: GET /api/v1/usage (plan, credits left, this key's counters). Spends nothing.

{
  "object": "answer",
  "id": "ans_...",
  "answer": "Probably not yet. ...",
  "agreement": { "level": "split", "agree": 3, "disagree": 1, "responded": 4, "consulted": 4, "summary": "3 of 4 models agreed ..." },
  "disagreement": {
    "headline": "The models split on timing",
    "topic": "...",
    "tie": false,
    "positions": [ { "stance": "...", "summary": "...", "models": ["..."], "count": 3, "of": 4 } ],
    "agreed": ["..."], "differ": ["..."],
    "why": { "recorded": true, "text": "..." }
  },
  "models": [ { "name": "...", "provider": "...", "status": "responded" } ],
  "sources": [ { "kind": "...", "title": "...", "url": null } ],
  "source_check": null,
  "usage": { "credits": 12 },
  "created": "2026-10-04T12:00:00.000Z"
}

Read sources carefully. They are what the models cited. Keplar has not checked them unless source_check is present, and agreement between models is not proof.

Set "stream": true or send Accept: text/event-stream. Events arrive as progress, delta (a live preview), answer (the result), then done or error. Refusals such as limits come back as a normal HTTP status before any event.

curl -N https://keplar.one/api/v1/ask \
  -H "Authorization: Bearer $KEPLAR_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"question": "Is it safe to store JWTs in localStorage?", "stream": true}'

event: progress
data: {"stage":"choosing_models"}

event: progress
data: {"stage":"comparing"}

event: delta
data: {"text":"Short answer: ..."}

event: answer
data: { ...the same object as the JSON response... }

event: done
data: {}

Create a widget in Account → Embed Ask. You get a public id (emb_…) and a snippet. Paste it where the box should appear. Questions are answered by Keplar and spend the publisher's credits. Free publishers always run on OpenRouter :free models, and paid publishers choose free models or their plan's models for each widget.

<script src="https://keplar.one/embed.js"
  data-keplar-embed="emb_YOUR_WIDGET_ID"
  data-theme="dark"
  data-size="standard"
  data-accent="graphite"
  async></script>

• Allowlist. Add up to 10 origins such as https://docs.example.com. Other sites get a refusal from the server, not only a hidden box.
• Rate limits. Each widget has its own per-minute and per-day limit by plan, plus a per-visitor limit.
• Revoke any time. A revoked widget stops answering on the next question.
• The id is public by design. It can only ask through that widget's limits. Never put an API key in a page.

A team is its own workspace. Projects, chats and sites created while you work in it belong to the team, not to anyone personally. Switch between personal and team workspaces from the account menu in the app. Up to 10 members, with one shared pool of 13,500 credits a month. Business plan.

Invites are drafts until you send them, by email or as a one-time link that expires in 7 days and works only for the invited address. The server checks every action against your role on every request. Open Team workspace.

Your link looks like keplar.one/?ref=yourcode. When a friend opens it, creates a new account and verifies their email, you get 500 and they get 500 usage credits. Credits only, never cash. Up to 5 rewarded friends a month and 20 in total. Self-referrals and throwaway addresses earn nothing.

Copy your link · Referral terms

Every error has the shape {"error": {"code", "message", "retry_after"?}}. Widget errors use the same shape.

• Keys are stored only as SHA-256 hashes and shown once. Revoking a key takes effect on the next request.
• The API reads the Authorization header only. Browser cookies are never accepted, so a key can't use a session and a session can't call v1.
• API and widget calls are stateless. They don't read or write memory, save chats, or train on your questions. We keep per-key and per-widget counters so you can see usage.
• Exporting your account includes your keys and widgets, but never a secret. Deleting your account removes them. See the privacy policy.

• Team-level API keys, per-key credit caps, webhooks and an official SDK.
• Separate usage-based billing for API and widget traffic. Today it draws on your plan's credits.
• Tool use, file or image input, and Create through the API. No stated uptime figure during the beta.

Need one of these? Tell us. The machine-readable spec is at /api/v1/openapi.json (OpenAPI 3.1).