WZRD — AI Model Velocity Oracle for Agents

Which AI model should your agent use right now? Real-time adoption signals across 100+ LLMs. HuggingFace, GitHub, OpenRouter, ArtificialAnalysis — updated every 5 minutes. Dynamic model routing for autonomous agents. Agents earn CCM tokens on Solana.

🔭 Building agents that pay for APIs?

TWZRD's active product is agent trust + receipts for the x402 economy — score any resource before you pay, then get a signed, offline-verifiable receipt. It's live on Solana mainnet.

pip install twzrd-agent-intel

• Live API: https://intel.twzrd.xyz · Leaderboard: https://twzrd.xyz/leaderboard
• Free preflight: POST https://intel.twzrd.xyz/v1/intel/preflight
• Paid trust receipt: GET https://intel.twzrd.xyz/v1/intel/trust/{pubkey} (0.05 USDC, x402)

The model-velocity oracle below still works and is unchanged.

Quickstart (Python)

pip install wzrd-client

import wzrd

model = wzrd.pick("code")           # best model right now (changes daily)
details = wzrd.pick_details("code") # score, trend, confidence
top_5 = wzrd.shortlist("code", 5)   # ranked list

No API key. No account. Works immediately.

Preflight + Receipts (recommended before any paid inference or earning): Vet the seller wallet you're about to pay for the inference provider.

import httpx

# Free Readiness Card (preflight) — vet a resource/provider (and optionally the
# seller wallet you'd pay) before you spend.
resp = httpx.post(
    "https://intel.twzrd.xyz/v1/intel/preflight",
    json={
        "resource_name": "gemma-4-26B-A4B-it",
        "price_usdc": 0.05,
        "agent_intent": "inference before paying",
        # "seller_wallet": "<provider pubkey>",  # arms the settle-gate on a counterparty
    },
).json()

card = resp["readiness_card"]  # fields are nested under "readiness_card"
print(card["decision"], card["trust_score"], card["can_spend"], card["caveats"])
# decision == "block" -> do not pay.

# If you decide to proceed, fetch the paid trust intel + signed V5 receipt:
# GET /v1/intel/trust/{pubkey}?seller_wallet=<counterparty>  (0.05 USDC, x402)

Verify the receipt offline — close the loop by trusting nothing but the bytes and TWZRD's published key (twzrd-receipt-verifier):

npx twzrd-receipt-verifier resp.json --pubkey 9V6Pn19kiUA5Rn6JpQfNduanvGt2aXGwsarosNfa2Ldf

Via MCP (first-class for agents):

• twzrd.get_readiness_card_tool (resource_name, price_usdc, agent_intent)
• Example with gemma-4-26B-A4B-it returns default_no_data (trust_score ~45, warn, can_spend: false) when no prior V5 data.

pip install twzrd-agent-intel  # for direct MCP/HTTP trust layer

Task types: code, chat, reasoning, math, multilingual — or any string.

Before acting on any velocity signal or paying: call the Readiness Card to vet the seller wallet you're about to pay. Attach V5 receipt on every paid response. This is how you get on the paid graph.

Quickstart (TypeScript)

npm install @wzrd_sol/sdk

import { bestModel } from '@wzrd_sol/sdk';

const picks = await bestModel({ task: 'code', budget: 'micro' });
console.log(picks[0].model_id);

Framework plugins on npm:

• @wzrd_sol/eliza-plugin — ElizaOS
• @wzrd_sol/solana-agent-plugin — Solana Agent Kit
• @wzrd_sol/goat-plugin — GOAT SDK

Why WZRD

• Trust before you pay — Free Readiness Card (preflight) scores the seller wallet for any model/provider before you spend. Attach a signed V5 receipt (0.05 USDC via x402) on every inference. Use the MCP tool twzrd.get_readiness_card_tool (or POST https://intel.twzrd.xyz/v1/intel/preflight). Vet the seller wallet you're about to pay before acting on any velocity signal. This is how you get on the paid graph. Live on Solana mainnet with twzrd-agent-intel.
• Save money — models change weekly. The one you hardcoded is probably 10x more expensive than the trending alternative.
• Better results — momentum is a leading indicator. Models gaining adoption fast are usually improving fast.
• Trustless on-chain oracles — 9 Switchboard feeds on Solana mainnet. Verify any signal independently.
• Get paid — agents that report inference results earn CCM tokens through a gasless relay.

Use Cases

• Autonomous agents that always pick the fastest/cheapest/best model
• Multi-agent orchestration (CrewAI, LangGraph, AutoGen, Eliza)
• MCP clients (Claude Code, Cursor) — 17 tools via the live MCP server at intel.twzrd.xyz/mcp
• On-chain protocols that need verifiable model selection data

Earn CCM Tokens (optional)

Agents that report which model they picked — and what happened — earn CCM on Solana. The usage data improves the oracle, so WZRD pays for it.

wzrd.run_loop()
# authenticates → picks models → runs inference → reports → claims CCM

What you need: Nothing. The client auto-generates a Solana keypair at ~/.config/solana/wzrd-agent.json on first run. Claims are gasless — no SOL needed.

CLI equivalent:

wzrd verified-demo           # one-shot verified earn-path demo
wzrd run                     # earn loop: pick → report → claim

Full Python API

Candidate-Aware Routing

Constrain picks to models you actually have access to:

model = wzrd.pick(
    "code",
    candidates=[
        "openrouter/qwen/qwen3.5-9b",
        "openrouter/qwen/qwen3.5-35b-a3b",
        "anthropic/claude-sonnet-4.6",
    ],
)

Agent Auth

agent = wzrd.WZRDAgent.from_env()
session = agent.authenticate()
receipt = agent.report_pick(choice, quality_score=0.9, latency_ms=1200)
status = agent.earned()

Keypair loading: ~/.config/solana/id.json, WZRD_AGENT_KEYPAIR_PATH, WZRD_AGENT_KEYPAIR (base58 or JSON byte array).

Public REST API

GET https://api.twzrd.xyz/v1/signals/momentum

Full OpenAPI spec: api.twzrd.xyz/openapi.json

On-Chain Oracles

9 Switchboard pull feeds on Solana mainnet (7 velocity + 2 price):

Full registry: wzrd.oracle.list_feeds()

On-Chain Identifiers

Links

• twzrd.xyz/start — onboarding
• twzrd.xyz/feed — live velocity feed
• MCP guide — connect to Claude Code / Cursor
• PyPI — Python package
• npm — TypeScript SDK
• twzrd-agent-intel — trust + receipts layer (preflight, V5 receipts, MCP)
• twzrd-receipt-verifier — standalone offline V5 receipt verifier

Environment Variables

License

MIT