your agent stopsrepeating mistakes

Infrastructure for AI agents to learn and grow. GreenCube sits between your agent and the LLM, tracks what works, and stops repeating mistakes.

For developers using OpenAI, OpenClaw, Ollama, LM Studio, or any OpenAI-compatible API.

How it works
A proxy that grows.
Point your agent at GreenCube instead of OpenAI. Every request passes through. Every response teaches it something.
> Your agent sends a request Claude Code, Cursor, OpenClaw, any agent
GreenCube intercepts it injects knowledge, state, preferences
Forwards to your LLM OpenAI, Ollama, OpenRouter, LM Studio
Response comes back extracts facts, scores quality, updates competence
Agent gets smarter next task uses everything it learned
Install
Three commands. That's it.
maccurl -fsSL https://greencube.world/install.sh | bash
winpowershell -c "irm https://greencube.world/install.ps1 | iex"
thenexport OPENAI_API_BASE=http://localhost:9000/v1
Run your agent normally. Then check what it learned:
gcsee your agent's brain
curl localhost:9000/statusone-line summary
curl localhost:9000/logrecent activity
curl -X POST localhost:9000/feedback -d '{"rating":"down"}'thumbs-down a bad response. your agent won't make that mistake again.
What happens after you install
It gets smarter every day.
Day 1
Your agent learns 12 facts from your first few tasks. Type gc to see them.
Day 3
It caught a mistake you made last week before you made it again. You didn't ask it to.
Day 5
Competence bars show python 87%, css 43%. Your agent knows what it's good at.
Day 7
Background reviews run when idle. Daily audits show what the agent learned.
Day 10
If your agent keeps struggling in one area, GreenCube spins up a focused sub-agent for that domain and routes matching tasks to it automatically.
Under the hood
5 systems. One runtime.
Runs locally. 145 tests. SQLite database you own.

Memory

Extracts facts from every task. Injects relevant ones into future requests. Decays unused entries over time.

Self-verification

Scores every response 1-5 against the original request. Low scores drop competence and trigger retries.

Competence tracking

Per-domain confidence from real outcomes. Knows python 87%, css 43%. Routes tasks based on actual success rates.

Mistake prevention

Stores corrections from thumbs-down feedback. Catches matching patterns in future tasks before they fail.

Auto-routing

When the agent keeps struggling in a domain, spins up a focused sub-agent for that domain. No user action required.

Compatible with
Works with: OpenAI, Ollama, LM Studio, OpenRouter, OpenClaw, LangChain, CrewAI, and any OpenAI-compatible API.
Claude: works through OpenClaw or OpenRouter. Native Anthropic API support coming soon.
Give your agent a brain.
Free, open source, runs locally. No cloud. No telemetry. Your data stays yours.
View on GitHub
MIT license · written in Rust · 145 tests