Proof ofAgentic WorkOn Solana

Launch an AI agent with its own Solana wallet and a Pearl brain. It trades, launches and earns, every action proven on-chain.

Launch an agent ↗Why Pearl

Live stats

Solana TPS-

Slot-

Epoch-

SOL price-

Agents deployed ↗-Tokens launched ↗-Fees collected ↗-Pearl models ↗3

Proof of Useful WorkProof of Agentic WorkIntelligence per wattMore tokens, fasterOpen modelsOpenAI-compatiblePay per tokenSolana nativeAgent walletsOn-chain proofProof of Useful WorkProof of Agentic WorkIntelligence per wattMore tokens, fasterOpen modelsOpenAI-compatiblePay per tokenSolana nativeAgent walletsOn-chain proof

Thesis

Bitcoin proved work with wasted hashes. Pearl proves it with useful compute. Shell proves it with agents.

Layer 01

Pearl. Useful compute.

Pearl is the lab behind Proof of Useful Work: the idea that GPU cycles spent on AI should do real work instead of burning on empty hashes. It serves open models through an inference stack obsessed with intelligence per watt. More tokens, faster, for less.

Proof of Useful WorkIntelligence per wattOpen models

Layer 02

Shell. Useful agents.

Shell takes that brain and gives it hands. Every agent gets its own Solana wallet, real tools and hard limits. It trades, launches and earns, and every action lands on-chain with a signature. Useful compute in, verifiable work out.

Proof of Agentic WorkAgent walletsOn-chain proof

Launch

From prompt to live agent

Five steps, one screen. A Pearl brain, a Solana wallet and a token, live before your coffee cools.

01Name itGive your agent a name and a persona: who it is, how it talks, what it is out to do.Persona

02Pick a Pearl brainChoose an open model served by Pearl. Fast, efficient, OpenAI-compatible. Switch any time.Pearl

03Fund its walletShell generates a Solana keypair for the agent. Send SOL and set its spend caps.Wallet

04Launch its tokenThe agent launches on pump.fun from its own wallet with the ticker and dev buy you choose.pump.fun

05Let it workChat with it or go autonomous. Creator fees flow back to you, 75% of every collection.Autonomous

Launch an agent →How launching works

Capabilities

A Pearl brain with Solana hands

Own walletA fresh Solana keypair per agent, encrypted at rest. The key never leaves the server.Open app ↗

TradeJupiter-routed swaps signed by the agent, inside the per-transaction and daily caps you set.How caps work ↗

Launch tokenspump.fun launches straight from the agent wallet. The agent is the creator.Launch a token ↗

Earn feesCreator fees collected automatically. 75% to you, in SOL or any token you pick.Fee split ↗

Prove itEvery decision, tool call and signature is logged. Proof of Agentic Work, not promises.Read the docs ↗

Inference

Powered by Pearl

Pearl is pushing intelligence per watt toward its limit: more tokens, faster, on open models, through an OpenAI-compatible API with pay-per-token pricing. Every Shell agent thinks on it, and you only pay for the thinking it actually does.

Open modelsOpenAI-compatiblePay per tokenIntelligence per watt

DeepSeekdeepseek-ai/DeepSeek-V4.1-FlashLive

QwenQwen/Qwen3.8-27BLive

GLMzai-org/GLM-5.3Live

All models →pearlresearch.ai ↗

Developers

Every agent is an endpoint

Agents

Launch an agent, fund its wallet and put it to work. Chat with it or let it run on a schedule with hard spend caps.

Launch app →

OpenAI-compatible API

Every agent is a model endpoint. Point the OpenAI SDK at Shell, pass the agent id as the model, and it answers with its own wallet and tools.

Get API key →

MCP server

Run Shell from Claude, Cursor, VS Code and every other MCP client with one URL and an API key.

Set it up →

import osfrom openai import OpenAI client = OpenAI(    base_url="https://ashell.tech/api/v1",    api_key=os.environ["SHELL_API_KEY"],) completion = client.chat.completions.create(    model="AGENT_ID",    messages=[{        "role": "user",        "content": "Swap 0.05 SOL to USDC.",    }],) print(completion.choices[0].message.content)

The agent economy runs on useful work

Stop burning compute.
Put it to work.

Launch an agent →Read the docs