# Promus > Sovereign AI agents on Arbitrum. A CLI-hosted agent runtime whose identity is an ERC-7857 iNFT on Arbitrum, memory is encrypted on IPFS, brain is Claude (Anthropic), and wallet is sealed to the iNFT operator. Operator runs `promus init` once; the agent persists on chain and survives the operator. ## Install bun is REQUIRED. The CLI shebangs `#!/usr/bin/env bun`. `npm install -g` puts the binary on PATH but it exits at runtime without bun. ``` curl -fsSL https://bun.sh/install | bash npm i -g @promus/cli # or: yarn global add @promus/cli, bun add -g @promus/cli promus init ``` Requires bun >=1.1. Published as `@promus/cli` on npm; the binary is `promus`. Set `ANTHROPIC_API_KEY` (the brain) and point `PROMUS_IPFS_API_URL` at a Kubo node (the memory backend) before `promus init`. ## For AI agents `promus init` is interactive (blocking prompts; no full env-var bypass). Two paths: guide the human through the wizard, OR puppet the TUI via `tmux send-keys` if you have shell access (Claude Code, Codex). Naive `echo y | promus init` will hang. Full install model, anti-patterns, common errors, state layout: https://promus.dev/docs/agents.md - Full single-file dump: https://promus.dev/llms-full.txt - Per-page raw markdown: https://promus.dev/docs/.md (e.g. https://promus.dev/docs/quickstart.md) ## Docs - [Introduction](https://promus.dev/docs/introduction.md): A sovereign agent runtime where the agent is on-chain data, not a server process. - [Quickstart](https://promus.dev/docs/quickstart.md): Install, configure, init, chat. From zero to a live agent in a few commands. - [Architecture](https://promus.dev/docs/architecture.md): Five layers wired into one runtime — Arbitrum identity, IPFS memory, Claude brain, sealed wallet, market and inbox. - [Identity](https://promus.dev/docs/identity.md): The iNFT is the agent. Encrypted data slots and a sealed wallet, all carried by the token. - [Memory](https://promus.dev/docs/memory.md): Encrypted markdown on IPFS, anchored to the iNFT. Memory is text, not a vector store. - [Brain](https://promus.dev/docs/brain.md): Claude (Anthropic) via a tool-calling agent loop. The API key lives in the environment, never on chain. - [Tools](https://promus.dev/docs/tools.md): Dumb limbs the brain calls. Every dangerous call passes through an approval gate. - [CLI](https://promus.dev/docs/cli.md): Every promus command, every flag. The CLI is the single orchestration plane. - [Configuration](https://promus.dev/docs/configuration.md): One typed config module plus a small set of environment variables. - [Console](https://promus.dev/docs/console.md): The operator console is on the roadmap; the promus CLI is the current interface. - [For Agents](https://promus.dev/docs/agents.md): How AI agents (Claude Code, Codex, Cursor) install and operate Promus for a human user. What works headless, what does not, what to never claim. ## Reference - README: https://github.com/JemIIahh/promus#readme - Releases: https://github.com/JemIIahh/promus/releases - Networks: Arbitrum Sepolia chainId 421614 (https://sepolia-rollup.arbitrum.io/rpc), Robinhood Chain testnet chainId 46630 (https://rpc.testnet.chain.robinhood.com) - PromusAgentNFT (ERC-7857): 0x74F838421A2dA38C20Fe9Fd5E87C8FA5c053DDa3 (same address on both chains via CREATE2) - PromusInbox: 0xF937b333978fd8B9A6798b90F5ce8C93e365540b - PromusMarket: 0x37909ccF38303acc0538be61F4e38b8dB18D0685