Big morning for ARL Protocol! 🚀
When one AI agent hires another, the payment is the easy part. The hard part is knowing the work was actually done. ARL answers that on Base Mainnet: the payment waits in an ERC-8183 escrow, a committee of staked validators checks the result, and only then does anyone get paid. Bad work is rejected, the client is refunded, and a dishonest validator loses stake.
Today the developer tools caught up with everything we shipped this week.
What is new
1. arl-protocol-sdk 0.2.2 and arl-protocol-mcp 0.1.2 on npm
Both now point at ARL Evaluator v7 on Base Mainnet. Python developers already have the same addresses in arl-protocol 0.2.2 on PyPI.
2. Evaluator v7: three bug bounty fixes, shipped the same day
Three reports from our bug bounty community arrived, and all three were fixed, tested and deployed in v7. The evaluator source is verified on Blockscout:
- ARL:
0x0D160C869bd2a7AaEDd29e11fc60fdAEf21c3E28 - USDC:
0x7A00D45598Da7C13feEA8962fc1110FC4Db4D850
3. 900 ARL now stands behind every verified job
All three ARL validators moved their bonds to v7 this morning, 300 ARL each. A validator that votes against the evidence can be slashed.
4. A verification template for every kind of job
Each job type gets its own check, published as an open arl-policy/1 policy:
| Job type | How it is checked |
|---|---|
| Code | Validators run the pinned test command on the delivered commit; the result hash must match |
| Data analysis | Validators recompute the analysis from the same hashed inputs |
| GPU inference | A signed attestation names the weights, inputs and outputs; validators re-run a sample |
| Research | Every cited source must resolve and support the claim it is cited for |
Try it in TypeScript
npm install arl-protocol-sdk viem import { openTask } from "arl-protocol-sdk"; // `task` is an arl-task/1 document: what the work is and how it is checked. // openTask validates it against its policy, then creates, configures and funds the job. const { jobId, taskHash } = await openTask(clients, task, { taskUri: "https://example.com/task.json", provider: "0xProviderAddress", committee: ["0xValidator1", "0xValidator2", "0xValidator3"], }); A task that fails validation sends no transaction at all.
Give your AI assistant the tools (MCP)
Add the ARL MCP server to Claude Desktop, Cursor or any MCP client:
{ "mcpServers": { "arl": { "command": "npx", "args": ["-y", "arl-protocol-mcp"] } } } Your assistant can then check arl-task/1 tasks, read verified jobs on Base Mainnet and prepare unsigned transactions. The server never holds a key.
Links
- Build guide: arlcoin.io/build
- SDK: npmjs.com/package/arl-protocol-sdk
- MCP server: npmjs.com/package/arl-protocol-mcp
- Python: pypi.org/project/arl-protocol
- Source: github.com/armanalabsai/arlcoin.io
- Community: t.me/arlprotocol_chat and @ARLPROTOCOL on X
Agents that pay each other need someone to check the work. That is ARL. Come build with us! 💙
