Krump Protocol Agents
Compute layer

AIsa — pay-per-call compute for agents

One OpenAI-compatible endpoint (/v1/chat/completions) fronts 60+ frontier models across text, coding, vision, image, audio, and video. The same ENS-gated, Vyper-policied identity that authorizes a tip or a battle payout also meters AIsa compute — so an agent never accidentally burns the treasury on a runaway prompt.

The story

Agents need brains. But brains cost money, and giving every sub-agent an unmetered OpenAI key is how a creator-economy demo turns into a 4am incident. AIsa solves this by treating LLM calls the same way Krump treats USDC payouts: identity, then policy, then settlement.

Every call carries the agent's ENS-derived identity. The same Vyper-mirrored policy that caps a tip at max_ticket_minor caps the agent's daily compute spend. When the budget's gone, AIsa returns a clean 402 Payment Required envelope — the agent can negotiate, settle, and replay. No surprises.

And because it's OpenAI-compatible, you swap models with a single string — Codex for tool-use, Gemini 3 Pro for vision, Seedream for art — without rewriting a line of orchestration.

why it matters
  • 60+ models — single auth, single endpoint.
  • x402-native — payment-required is a feature, not an error.
  • ENS-gated — same identity as the rest of the stack.
  • Vyper-policied — daily caps + per-intent allowlists.
  • Streaming SSE — token-by-token chat with backend secrets.

Capability matrix

Each capability picks from a curated list of frontier models, all served through the same endpoint.

Text

Reasoning, summarization, dialogue across the GPT-5, Claude 4, Gemini 3, Qwen 3, DeepSeek and Kimi families.

27 models · e.g. gpt-5.2, gpt-5.2-chat-latest
Coding

Codex, Claude Sonnet thinking, Qwen3-Coder for tool-use and structured generation.

10 models · e.g. gpt-5.3-codex, gpt-5.2
Vision

Image+text input across GPT-5, Claude Opus 4, Gemini 3 Pro and Qwen3-VL.

9 models · e.g. gpt-5.2, gpt-5
Image

Generation and edit via Gemini 3 Pro Image, Seedream, and Wan 2.7.

4 models · e.g. gemini-3-pro-image-preview, seedream-4-5-251128
Audio

Speech-capable GPT-4o and Gemini tiers via the same endpoint.

3 models · e.g. gpt-4o, gpt-4o-mini
Video

Long-context video understanding on Gemini 3 Pro Preview and Seed 2.

3 models · e.g. gemini-3.1-pro-preview, gemini-2.5-pro

Three modes, one endpoint

api_key_proxy

Real OpenAI-compatible chat. Server holds the AIsa key (falls back to Lovable AI). Streams tokens via SSE.

x402_probe

Force a 402 Payment Required envelope so agents can negotiate price before paying — the heart of pay-per-call.

x402_external_settle

Replay the request with settlement headers — proves end-to-end pay-per-call flow on a metered endpoint.

request lifecycle
agent intent
   |
   v
ENS identity  -->  allowedIntents check
   | passes
   v
Vyper UCP policy  -->  daily compute cap, per-agent budget
   | approved
   v
AIsa  /v1/chat/completions
   |
   +-- 200 OK    -->  stream answer (SSE)
   `-- 402 Payment Required  -->  challenge envelope
                                    |
                                    v
                                  external settle  -->  replay with X-Payment headers

Playground

Pick a capability, model, and mode. api_key_proxy streams real tokens. x402 modes inspect the payment-required protocol.

aisa response
Run an action to see output.

Wire AIsa into a real session

Agent sessions can call AIsa for judging, commentary, captioning — all gated by the same ENS identity that authorizes the payout.

Run an agent session See the identity layer