// open-source · MIT · self-hosted · crypto-native

Give it a goal.
It evolves to get there.

POLYROB is a self-hosted autonomous agent you run on your own machine. Give it a goal in plain language and it does the rest — planning the work, browsing the web, reading and writing files, running code, and recovering from its own errors until the goal is done. Switch on personal-agent mode and it becomes durably autonomous: it writes and curates its own skills, keeps a backlog of goals that survives restarts, wakes itself to follow up, and refines an evolving model of how you work.

Self-hosted, not a SaaS. There is no hosted version — your keys, your data, your control.

Run it in a few lines

$ pipx install "polyrob[all]" && polyrob init && polyrob chat

Bring a provider key (OpenAI / Anthropic / Gemini / …) and you're chatting in minutes. Set POLYROB_LOCAL=true to unlock the self-evolving + goal-seeking loops. Before the PyPI package is live you can install from source: pipx install "polyrob[all] @ git+https://github.com/theselfruleorg/polyrob". Full install guide →

What makes it different

🧬 It learns and evolves itself

Most agents forget everything when a task ends. POLYROB writes its own skills from what it learns, retires and revives them automatically, and refines an evolving self-identity — every self-modification quarantined and reviewed before it takes effect.

🎯 It pursues goals durably

Hand it a goal and walk away. A durable goal board (atomic claims, circuit breakers) survives process restarts — a normal agent loses in-flight work when it crashes; POLYROB picks it back up. Plus self-wake, cron, and least-privilege delegation.

🛡️ It's hard to hijack

Autonomy is only safe if untrusted input can't seize the wheel. Web pages, emails and tool output are framed as data, never instructions; sub-agents are least-privilege; and a 3-tier access model gates high-impact tools the moment a third party is in the loop.

H-MEM hierarchical memory

Not a flat log: findings are organized into semantic phases, consolidated by an LLM, and forgotten by importance (recency + relevance + frequency), not age. Cross-session recall is keyword by default, semantic vector optional — an implementation of H-MEM.

Crypto-native

A built-in agent wallet (policy-capped, audited), x402 pay-per-request in USDC on Base, Avalanche & IoTeX, and on-chain identity + reputation via ERC-8004 — the agent pays, charges, and carries verifiable trustless reputation. Market data via Polymarket & Hyperliquid.

Connects to everything

Browser automation, AnySite structured-data sources, MCP servers, code execution, documents, voice. Reaches you across chat surfaces — terminal, Telegram, WhatsApp, email — through one contract. Any of 6 LLM providers, switch mid-session.

Build on it

POLYROB is self-hosted — there's no hosted version. But the multi-tenant isolation, usage metering, credits and on-chain payment primitives are in the core, so if you want to build your own agent product on your own instance, you're not bolting a backend onto a toy.

Multi-tenant + metering

Per-tenant isolation, usage metering, credit balances and pay-per-request — serve many users and meter exactly what they spend.

Standard interfaces

A REST API, an OpenAI-compatible /v1 surface (drop-in for existing SDKs), and Google's A2A protocol so other agents can discover and use yours.

On-chain economy

ERC-8004 identity + reputation, A2A discovery and x402 payments make your agent a discoverable, payable, reputation-backed participant in an open agent economy.

Self-host the whole thing, free

Full engine, MIT✓ free forever
Your keys, your data, your machine✓ no hosted version
Goal board + autonomy loops✓ survives restarts
Security that's actually on by default✓ hard to hijack

See the full, dated comparison →