Show HN: Polpo – Build zero-human companies. Open source(github.com)
github.com
Show HN: Polpo – Build zero-human companies. Open source
https://github.com/lumea-labs/polpo
4 comments
The keychain-per-agent model is the right instinct, but the threat model gets complicated fast. When five agents share a keychain process, a prompt injection in agent A can exfiltrate credentials for agents B–E — one bad LLM output becomes a full company compromise.
The safer architecture is per-session credential scoping: each agent gets credentials minted for that specific task, revoked on completion. That's what we built Cyqle (https://cyqle.in) around — ephemeral sessions where the encryption key is destroyed on close, so if an agent misbehaves, the blast radius is bounded to that session.
The keychain needs the same isolation principle, but applied at the task boundary, not the agent boundary.
The safer architecture is per-session credential scoping: each agent gets credentials minted for that specific task, revoked on completion. That's what we built Cyqle (https://cyqle.in) around — ephemeral sessions where the encryption key is destroyed on close, so if an agent misbehaves, the blast radius is bounded to that session.
The keychain needs the same isolation principle, but applied at the task boundary, not the agent boundary.
The problem: you open Claude Code, give it a task, it does 80%. You fix the other 20%, open another chat for the next piece, copy context, retry when it drifts. Before you know it you're a full-time AI babysitter — 4 monitors, 12 terminals, zero confidence anything actually ships.
Polpo fixes this. You build an AI company: hire agents, give them roles, skills, and credentials stored in a keychain. They work as a team.
The key difference is quality control. Think of it like UFC judges, but for your agents' work:
For every task you define custom scoring criteria 3 independent LLM reviewers evaluate the output in parallel Median score vs threshold — below? The agent goes back to work automatically You don't even know it happened. Nothing ships broken. What else:
Crash-proof — detached processes, kill it, reboot, lose connection. Picks up where it left off Agents get real credentials, persistent browser sessions, real accounts — not sandboxed toys Reaches you on Slack, Telegram, or email only when it needs a human decision 22+ LLM providers including Ollama and local models — not an OpenAI wrapper Web UI, CLI, REST API, SSE streaming — all included Storage: flat files (default), SQLite, or Postgres One command:
npx polpo-ai
Claude Code gives you a developer. OpenClaw gives you an assistant. Polpo gives you a company. Marketing agency, dev team, customer support — all AI, on your laptop.
MIT licensed. TypeScript. Hono + React. No hosted platform, no waitlist — your keys, your rules.
GitHub: https://github.com/lumea-labs/polpo Docs: https://docs.polpo.sh Site: https://polpo.sh
Happy to go deep on the scoring architecture, the state machine, crash recovery, or anything else.