setup

Run your own agent

AgentCron is single-tenant by design. To use it for real, fork the repo, plug in your own keypair + Anthropic key, and run the agent locally. Your keys never leave your machine.

01this dashboard's agent

signed transactions you see come from this wallet

status
…
agent wallet
…
program
CJ2uqxr4XVD22a66hcpat524KUkAPTeZMmgM15tRFugC↗
cluster
devnet
02run it yourself

3 minutes from clone to live agent

1

clone the repo

git clone https://github.com/JkrishnaD/agent-corn
cd agent-corn
npm install
2

generate an agent keypair

Make a fresh keypair just for the agent. Don't reuse your main wallet.

solana-keygen new --outfile ~/.config/solana/agent-cron.json
solana-keygen pubkey ~/.config/solana/agent-cron.json
# copy that pubkey, fund it on devnet:
solana airdrop 2 <PUBKEY> --url devnet
3

export the secret as base58

The agent reads SOLANA_PRIVATE_KEY as a base58 string.

# one-liner: read JSON array, encode bs58
node -e "const fs=require('fs');const bs58=require('bs58');console.log(bs58.default.encode(Buffer.from(JSON.parse(fs.readFileSync(process.env.HOME+'/.config/solana/agent-cron.json')))))"
4

configure .env

Copy the template, fill in 3 keys.

cp .env.example .env
# Solana
SOLANA_NETWORK=devnet
HELIUS_API_KEY=<your helius key>
HELIUS_RPC_URL=https://devnet.helius-rpc.com/?api-key=<your helius key>
SOLANA_PRIVATE_KEY=<base58 of your agent keypair from step 3>

# LLM
ANTHROPIC_API_KEY=sk-ant-...

# MagicBlock router (devnet, public — no key needed)
MAGICBLOCK_RPC_URL=https://devnet.magicblock.app

# Pre-deployed program (devnet)
AGENT_CRON_PROGRAM_ID=CJ2uqxr4XVD22a66hcpat524KUkAPTeZMmgM15tRFugC

# Web
NEXT_PUBLIC_AGENT_API_URL=http://localhost:4000
NEXT_PUBLIC_AGENT_WS_URL=ws://localhost:4000
5

verify reachability

The smoke scripts confirm RPC, MagicBlock router, and the program lifecycle work end-to-end.

cd agent
npx tsx src/scripts/smoke-magicblock.ts   # 11 checks
npx tsx src/scripts/smoke-program.ts      # init → delegate → 3 ER ticks → commit
6

run the stack

# terminal 1
cd agent && npm run dev

# terminal 2
cd web && npm run dev

# open http://localhost:3000/dashboard
03trust model

who sees what

stays on your machine
  • SOLANA_PRIVATE_KEY (agent keypair)
  • ANTHROPIC_API_KEY
  • HELIUS_API_KEY
  • full LLM reasoning text + rule prompts
on chain (public)
  • AgentContext PDA per rule
  • execution counter, fired count, last confidence
  • sha256 of prompt + each rationale
  • tx signatures (transfer/swap)

Reasoning text never lands on chain. Anyone can re-hash your stored rationale to verify it matches what was recorded — agent can't retro-edit decisions.

04known limits

hackathon scope, not architecture

  • —Rules in memory only — agent restart loses rules. JSON-file persistence is trivial to add.
  • —Single-tenant: one agent keypair owns all PDAs. Multi-user needs keypair-per-rule auth.
  • —Monitor is RPC-poll (5s). Helius webhooks pluggable — interface is in monitor.ts.
  • —Only `transfer` action wired in dashboard. `swap` and `custom` work via direct POST /rules.
githubdevnet faucetopen dashboard