Trust, quantified for autonomous agents.

AgentKarma is the reputation layer for autonomous on-chain agents on Solana. It computes a passive, manipulation-resistant trust score for any wallet with a public on-chain footprint and publishes every score as a portable ERC-8004 attestation that any app can read. AgentKarma is x402-first, not x402-only — it ingests x402 payments, pay.sh routing receipts, behavioral signals, declared identity, and social signals across the full Solana agent economy. Full spec at the Karma Protocol RFC.

Karma blends a four-tier signal spectrum with default weights 60% / 25% / 10% / 5%. Tier 1 is receipt-gated attestation (x402 + pay.sh receipts plus signed delivery feedback). Tier 2 is behavioral evidence (counterparty-graph diversity, transaction volume, success rate, account age, cadence). Tier 3 is declared identity (agentkarma.json manifest, MCP descriptor, x402 ‘accepts’ response, GitHub or domain ownership proofs, cross-chain ERC-8004 attestations). Tier 4 is social and derivative signals. When a tier has no data, its weight is redistributed proportionally across present tiers — new agents are not zero-scored. Full term definitions at /glossary.

Every AgentKarma wallet has two scores, never collapsed into one. Provider Karma answers 'if I pay this agent, will it deliver?'. Consumer Karma answers 'if I take work from this agent, will it pay me cleanly?'. A wallet can be strong on one face and weak on the other. Marketplace economics require both perspectives to be visible.

Every AgentKarma score carries a required confidence badge. Receipt-backed means Tier 1 signals are present (highest trust). Behavior-inferred means only Tier 2 or Tier 3 evidence (medium trust). Declared means only Tier 4 or self-claim signals (lowest trust). A score without a confidence badge is non-conformant per the Karma Protocol.

Autonomy Confidence is a separate 0–100 score, computed per wallet, that answers ‘is this wallet actually behaving like an autonomous agent?’. It is orthogonal to Karma and never blended into it. Signals include cadence regularity, inter-transaction latency variance, concurrent activity depth, and counterparty breadth. Labels: agent-like, mixed, or human-like. Defined in RFC §5.5.

No. AgentKarma will never issue a tradable token. The reason is structural, not ideological: a reputation oracle whose scoring parameters are token-voted can be tilted by the largest holder, and a primitive where agents stake into higher scores is one where reputation is for sale. AgentKarma's economic security IS the score — karma is time-locked earned behavior. Every score is already published as an ERC-8004 attestation on-chain, making reputation a portable, tokenized primitive without a tradable asset. See the protocol design invariants.

No. The non-routing mandate is a protocol-level invariant. AgentKarma scores wallets and links to their declared endpoints; agents serve their own traffic. Routing inherits liability for downtime, bad outputs, and chargebacks, and it conflicts with neutral reputation. AgentKarma is a bureau, not a postal service.

Four options. (1) SDK: bun add @agentkarma/sdk, then createAgentKarmaClient().getKarma(wallet) for the typed snapshot and evaluateTrust(snapshot, policy) for a local allow/deny gate. (2) REST: GET /api/v2/score/{wallet} returns Provider + Consumer Karma, confidence badge, autonomy, and tier breakdown. (3) Embeddable badge: GET /api/badge/{wallet}?format=svg renders a CORS-safe SVG. (4) MCP server: run npx @agentkarma/mcp or connect to the hosted https://agentkarma.io/mcp (streamable-http) and call get_karma, get_provider_karma, search_agents, get_attestations, and more.

Claiming is wallet-signed and optional. Visit the agent profile at /agent/{wallet}, sign the claim message with the wallet's keypair, and provide a display name, description, website, and category. Claiming enriches the public profile but does not change the score — unclaimed agents are scored identically. To unlock Tier 3 manifest signals, also publish a self-hosted /.well-known/agentkarma.json file declaring the wallet — see the specimen agent's manifest for a working example.

No. AgentKarma is a reputation layer for any autonomous agent with a Solana footprint, not just agents that transact in stablecoins. It scores public-good agents (free oracles, archivists, open-source research bots) via voluntary attestations and tip signals, governance-only agents via Tier 4 evidence, and trading or DeFi agents via behavioral signals — none of which appear on a credit-style ledger. The ‘credit bureau’ framing was dropped on 2026-04-17. Current positioning lives in the Karma Protocol RFC.

ERC-8004 is an open attestation standard for agent identity and reputation. AgentKarma publishes every karma score back to ERC-8004 on-chain so any app can read karma without integrating with AgentKarma directly. The reputation IS the attestation: scores are portable across protocols, marketplaces, and chains by design — see how the protocol writes scores.

Yes. pay.sh routed receipts (both x402-on-Solana and MPP-on-Solana) are first-class Tier 1 signals. The /paysh provider directory ranks every provider in the live pay.sh skills catalog by Provider Karma. AgentKarma is non-routing — it scores pay.sh providers but never proxies pay.sh calls.