MCP tells agents how

MCP (Anthropic, Nov 2024) standardized how AI agents discover and use tools — it's now adopted by Claude, ChatGPT, Google, Microsoft, and IBM. x402 (Coinbase, May 2025) standardized how agents pay -- an open standard now adopted across the industry with 100M+ payments processed. Neither has a trust layer. Aegis is that layer: bonded validation, on-chain reputation, and atomic revenue splits for 412K+ skills. Agent Aegis is the runtime — a single binary that wires MCP discovery, x402 payments, and Solana settlement into one agent loop.

Micropayments for AI skills need sub-cent transaction costs and sub-second finality. Solana delivers approximately $0.00025 per transaction at 400ms finality with 4,000+ TPS. Ethereum L1 costs $2-50 per transaction, which kills any micropayment model. We bridge to Ethereum via ERC-8004 for identity, but settlement happens on Solana because it's the only chain where this is economically viable.

Validators stake $AEGIS bonds to attest that a skill works as advertised. If their attestation is accurate, they earn 20% of every invocation fee. If they rubber-stamp garbage, their $AEGIS bond gets slashed. This creates economic skin-in-the-game for quality. The mechanism is backed by MIT CSAIL research proving binding reward transfers produce socially optimal equilibria.

HTTP 402 (Payment Required) has been a reserved status code since 1997. For 28 years, nobody activated it. In May 2025, Coinbase launched x402 — a protocol that lets any HTTP server charge for access via standard headers. The x402 open standard has been adopted by major infrastructure providers, with over 100M payments processed. Minimum payment: $0.001 (vs Stripe's $0.30 minimum). When an agent invokes an Aegis skill, the server responds 402. The agent pays USDC. Aegis swaps to $AEGIS via Jupiter and splits revenue atomically. Every invocation is a buy order.

ERC-8004 is the Agent Registration standard on Ethereum. It defines how AI agents register their identity, capabilities, and trust models. The spec explicitly states that incentives and slashing related to validation are outside the scope of the registry. That open slot is Aegis. We implement the crypto-economic trust model that ERC-8004 names but doesn't build.

Every skill invocation fee is split atomically in a single Solana transaction: 70% to the skill creator, 20% to the validator who attested quality, 9% to the $AEGIS protocol treasury, and 1% is burned permanently. This is enforced on-chain by the Solana program. No invoices, no net-30, no chargebacks.

$AEGIS is the protocol token built on Solana using Token-2022 extensions. Total supply is 1 billion. It has six utility mechanisms: validator staking, skill registration bonds, governance voting, fee discounts for holders, dispute resolution staking, and a built-in 1% transfer fee with 0.5% permanent burn. Every protocol action requires $AEGIS.

Hugging Face Spaces hosts 300K+ AI applications — image generators, text models, MCP servers, agent environments — all running for free with zero economic layer. No payments to creators, no quality bonds, no reputation system. Aegis indexes HF Spaces alongside GitHub and adds the missing infrastructure: x402 micropayments, bonded validation, and on-chain reputation. Think of HF Spaces as the unmonetized skill library. Aegis is the economic protocol that turns it into a marketplace.

Token-2022 enables protocol-level transfer fees without smart contract workarounds. Every time $AEGIS changes hands, 1% is automatically collected -- 0.5% goes to the protocol treasury and 0.5% is burned permanently. This is enforced at the token program level, not by a contract that can be upgraded or bypassed. Transfer hooks enforce the minimum bond requirement automatically -- if a creator tries to transfer bonded tokens below the threshold, the hook rejects the transaction at the protocol level.

Every skill runs inside a Deno-based permission sandbox. The runtime treats all code as untrusted by default. Skills must declare their permission requirements in SKILL.md (network endpoints, filesystem access, environment variables). The consumer's Agent Aegis runtime enforces these permissions at the process level. A malicious skill cannot exfiltrate data, spawn subprocesses, or read environment variables without explicit grants. Node.js and Bun have no equivalent security model.

Observation loops are replayable audit traces recorded during every skill invocation and validator challenge. They capture the exact input, sandbox execution logs, network calls, timing, and output. When a challenge is filed, the observation trace becomes evidence -- anyone can replay the exact sequence and verify the outcome. This is the agentic engineering pattern of deterministic scaffolding around non-deterministic AI, applied to economic accountability.

Solana programs are stateless -- all state lives in accounts passed to the program. Aegis uses separate Program Derived Accounts (PDAs) for skill metadata, bond vaults, and reputation scores. This means each data domain can be updated independently without locking the others, enabling parallel execution of multiple skill invocations without blocking. It also means each PDA can have its own access control -- bond vaults are controlled by the registry program, reputation PDAs by the reputation program.

Aegis skills are MCP-native, which means any agent platform that supports the Model Context Protocol can discover and invoke them. Claude Code and Claude Cowork connect via MCP servers -- skills appear as tools in the agent's context. Claude Remote Control lets developers monitor Aegis invocations from mobile while sessions run on their machine. OpenAI's Codex CLI and Codex App connect through the same MCP interface or via direct HTTP x402 calls. ChatGPT, Cursor, and Windsurf all support MCP tool discovery. Agent Aegis is our own runtime that wires everything together natively. The key insight: Aegis doesn't compete with these platforms -- it's the trust and payment layer they all use when invoking external skills.