Tool · software development / AI engineering tools
0-CODE
2026 · Tool · Node.js 18+, Bash / PowerShell
An autonomous development loop for Claude Code and OpenCode: the agent plans, writes, tests and commits on its own — and keeps working after the session ends.
0-CODE is an open-source framework (MIT, Node.js, zero external dependencies) that turns an AI agent into an independent worker: the loop state is kept in project files (STATE and ROADMAP), a single shared module makes the “continue or stop” decision, and the guard plugin is placed outside the working folder so that the agent cannot switch off its own restrictions. It works with Claude Code (a Stop hook) and with OpenCode (an external supervisor, any model via OpenRouter). The public release v1.20.0 came out on 18 September 2026; it was preceded by Python prototypes from late 2025 (claude-automation and three demos, MK1–MK3).
The task
Out of the box, an agent loses context between sessions, dies on network failures, spends the budget unnoticed, can delete its own safety rules and declares work finished on its word alone. 0-CODE makes the loop survive restarts, keeps its spending limited and makes it verifiable by data rather than by the agent’s claims.
What’s inside
- Loop state in files (docs/STATE.md, docs/ROADMAP.md): a new session picks up the work where the previous one broke off
- Two drivers with shared decision logic: a Stop hook for Claude Code and an external supervisor for OpenCode (Kimi, DeepSeek and any OpenRouter model)
- Minimal protection against irreversible commands (four rules); the guard plugin lives outside the working folder and is restored before every turn; arbiter / guard-only / allow-all / confirm modes
- Retries on network failures with increasing pauses, a loop lock (one owner per project), cost tracking with an optional budget ceiling
- A “no progress” detector, a registry of long-running tasks (running / overdue / finished / dead), decision records with at least three options
- A machine-readable register of audit findings: the /productionauditloop command runs until there are zero P0–P2 findings according to the data, not to a declaration
Numbers
- Open source, MIT, zero dependencies
- 2.5 hours of autonomous work — 6 commits for $10.25
Machine translation — being proofread.