Solana AI agents in 2026: agent wallets, MCP servers, and what's actually shipping

"AI x crypto on Solana" means at least four different things, mostly muddled together in pitch decks. There are agent-launched tokens (daos.fun-style), trader bots, AI-themed memecoins, and — the actually-shipping primitive layer — agent wallets, MCP servers, and Solana-aware agent frameworks.

This article is only the last category: the infrastructure an AI agent uses to actually do something on Solana. Past the narrative, here's what's real.

Agent wallets

An autonomous agent needs a key to sign with. Giving an LLM a raw private key is a security disaster — any prompt injection or misroute can drain the wallet. The 2025-2026 solution: wallets with explicit policies baked into the signing layer.

The two main providers:

• Crossmint Wallets — spending limits per period, whitelisted programs, delegated signers. Same SDK works for consumer wallets and agent wallets.
• Privy — server-side wallets with policy engines, originally consumer but increasingly used for agent scenarios.

// Crossmint agent wallet with a 1 SOL/day cap
import { createCrossmint, CrossmintWallets } from "@crossmint/wallets-sdk"

const wallets = CrossmintWallets.from(
  createCrossmint({ apiKey: process.env.CROSSMINT_SERVER_KEY! })
)

const wallet = await wallets.createWallet({
  chain: "solana",
  signer: { type: "api-key" },
  config: {
    policies: [
      { type: "spending-limit", value: "1 SOL", period: "day" },
      // future: { type: "program-allowlist", programs: [JUPITER_PROGRAM_ID] }
    ],
  },
})

// The agent code can now sendTransaction without you worrying
// about runaway losses
await wallet.sendTransaction({ transaction: someJupiterSwap })

The point isn't that the policy engine is unbreakable — it's that the agent's downside is bounded by a number you set explicitly, not by how well you trust the LLM.

Solana MCP

MCP (Model Context Protocol) is Anthropic's standard for exposing tools to an LLM. A Solana MCP server wraps Solana RPC, wallet adapters, Jupiter swaps, etc. as MCP tools that Claude (or any MCP client) can call directly.

The reference implementation is SendAI's Solana Agent Kit — Eliza-compatible plus a standalone MCP server mode. Helius also publishes an MCP server, as does Crossmint.

What an MCP-equipped LLM can do, out of the box:

• Check any wallet's SOL or SPL balance
• Quote a swap via Jupiter
• Execute a swap (using a policy-gated wallet)
• Look up a token by mint address or symbol
• Read a Helius transaction parsed history
• Look up an SNS domain or resolve a Helium hotspot

Plugging this into a Claude Desktop config is two lines of JSON. Plugging it into a custom agent loop is similar. The skill bar for "agent that does Solana things" dropped from "build your own SDK adapters" to "import an MCP server" in about six months.

Agent frameworks

Three main frameworks dominate the Solana agent stack right now:

• Solana Agent Kit (SendAI) — TypeScript-first. Eliza adapter, MCP server, direct programmatic API. Most active repo in the space.
• Eliza — broader Web3 agent framework, supports Solana via plugins. Originally Twitter-bot oriented; now used more for general autonomous agents.
• Helius Agent Tools — opinionated wrappers on top of Helius' RPC and DAS APIs. Less framework, more bundled tools.

The narrative vs the reality

Most of what's called "AI on Solana" isn't agents doing anything autonomous. It's:

• AI-themed memecoins — branding. Real token, no agent.
• Agent-launched tokens (daos.fun / virtuals) — a token launched in an agent's name. The agent is mostly a marketing wrapper around a tradeable token.
• AI trading bots — algorithmic trading. The "AI" varies from real ML to scripted rules.
• Agentic commerce — actual category. Agents execute payments, swaps, subscriptions on the user's behalf. Crossmint is building heavily here.

The infrastructure layer (wallets, MCP, kits) is real and ship-ready. The application layer is still mostly hype, with a handful of credible use cases (commerce, trading, on-chain operations) actually emerging from the noise.

What to actually build

Concrete agent projects that are working in 2026:

1. A swap agent — "buy me $50 of JTO when the price drops 5%." Crossmint wallet + Jupiter MCP + a price trigger. A weekend of work.
2. A treasury agent — auto-rebalance a multi-token treasury weekly. Spend policy keeps it safe.
3. A research agent — pure read, no signing. Pull Helius transaction history for a wallet, summarise PnL, post to Discord.
4. A claim agent — auto-claim staking rewards, airdrops, etc., when conditions are met.

References

• sendaifun/solana-agent-kit
• Crossmint — Agent wallets
• Model Context Protocol

The agent narrative outran the infrastructure for most of 2024. It's the opposite now — the primitives are quietly excellent, the applications are still catching up. Good time to build.