# AGNT Models
Base: https://models.agnt.gg/models/v1
Spec: https://models.agnt.gg/openapi.json
Docs: https://models.agnt.gg/docs.html
OpenAI-compatible. Set base_url to the Base above, api_key to your Models key, model to agnt-flash. Streaming, tool calls, JSON output and vision are supported. Context 1,000,000 tokens; up to 65,536 output tokens.
POST /agents creates a scoped agent account. Save its one-time secret (mdl_…). Use Authorization: Bearer <secret>.
POST /chat/completions runs the model. Send an Idempotency-Key; reuse it on an uncertain retry to get the receipt instead of a second run. Response adds requestId and usage {inputTokens, cachedInputTokens, outputTokens, credits, includedCredits, chargedMicroUSD}. When streaming, the final data frame before [DONE] carries usage and requestId.
POST /quote with the same body returns maximumCredits and maximumChargeMicroUSD before anything runs.
Credits: one credit = one cached input token; a new input token = 50 credits; an output token = 200 credits. $0.025 per million credits. Included credits are drawn first, then prepaid cash up to maxChargeMicroUSD. You are never charged past that cap. maxChargeMicroUSD defaults 0 (included-only).
Pay as you go with $10 minimum prepaid credit, or POST /hosting/purchase with an Idempotency-Key and tier starter|pro|business to activate a standalone monthly plan from that credit. No AGNT subscription required. Starter $5 = 230M credits; Pro $15 = 700M; Business $39 = 1,800M. Extra credits bill at $0.025 per million. Higher tier mid-month upgrades in place for the difference; lower tier waits for renewal. No rollover.
Funding: Stripe claim link or x402 v2 USDC on Base via POST /funding/x402. Do not sign without explicit owner funding authorization.
Limits: 120 requests/minute per account, plus a concurrent-request limit; 4 MB body; 512 messages; 256 tools; five-minute request timeout. Twenty failures in an hour pauses requests. Failed requests are not charged.
Prompts and replies are not stored and are never used for training; receipts hold token counts and charges only. Never put secrets in prompts.