Curated models

Klunq Proxy does not expose the whole market. It fronts a small, tested set behind an OpenAI-compatible list. During the public beta that list is a single shared free endpoint — paid models with real pricing arrive with monetization.

Beta: free-ai (shared free endpoint)

• Public name: free-ai — what clients send (data-model="free-ai").
• Behind the name sits an OpenRouter free-model endpoint, shared by all users: roughly 50 requests/day in total. When the daily pool is exhausted, inference returns rate-limit errors until it resets — this is an upstream limit, not a bug in your key or budget.
• Cost resolves to $0.0 per call, so budget checks are vacuous during beta (they cannot trip on zero spend) while v-keys, page-lock and spend tracking keep working normally. Tracked spend reads $0.0 everywhere.

List available models (with any valid key):

curl -H "Authorization: Bearer $VKEY" "$PROXY_URL/v1/models"
curl -H "Authorization: Bearer $TEAM_MGMT_KEY" "$PROXY_URL/v1/models"

Why small models are enough

The Klunq Widget is a page-agent: read content, click, fill forms, answer from the visible page. It does not need multi-step coding or heavy reasoning.

• Small ~30B open models answer “what is this page about?”, “fill checkout”, “compare plans” reliably.
• Pricing is 10–100× cheaper than frontier multi-hundred-bilion parameter models.
• Overkill models add latency and cost with little UX gain for this task.

Klunq tests each model against page tasks before adding it. The proxy will only list models that pass that bar.

What changes next

• Add more open models in the same weight class (names TBA).
• Real per-token prices per model.
• No widget change required — just data-model versus /v1/models.

Choosing a model for a v-key

Either leave models empty (all allowed) or restrict:

curl -X POST "$PROXY_URL/key/generate" \
  -H "Authorization: Bearer $TEAM_MGMT_KEY" \
  -H "Content-Type: application/json" \
  -d '{"key_alias":"landing","models":["free-ai"],"max_budget":2}'

The proxy will reject completions that request a non-allowed model.

Next: Integration — widget → proxy · Budgets