[E] EvAgent
Sign In
Self-evolving Local-first

EVA - Evolving Agent

A local-first AI agent that grows its own capabilities, discovers new skills, writes code, tests itself, and evolves, all on your own hardware.

v1.22.0 35 Skills Nemotron-3B 64k
Quick Install View Documentation →
EVA — Evolving Agent
How It Works

From user request to evolved capability in 7 steps.

01

User Request

A command or query triggers the pipeline.

02

Gap Detect

Analyzes if the request exceeds current capabilities.

03

Tier 1 Search

Search ecosystem and local repositories.

04

Tier 2 Synthesis

Code-generation of new skills when no match is found.

05

Critic Verify

Ephemeral replica reviews code for quality and safety.

06

User Confirm

Human confirmation gates the resolution.

07

Merge & Learn

Skill is merged into the ecosystem.

Anchored to real user failures • Human confirmation = reward signal • Trajectory fine-tuning on success
Architecture

The full system, across 22 ADRs.

The system is designed around a modular, tool-first philosophy. Every component is replaceable, every skill is independently testable, and the whole stack runs on consumer hardware.

Tool-First Pipeline

The model always has access to tools. Skills and routines are meta-tools in a native tool-calling loop. No routing-first triage.

ADR-022

Named Model Slots

Multi-slot architecture: primary (Nemotron), audio (Gemma 4), draft, embed — each independently managed with LRU eviction.

ADR-018

Multi-Provider Routing

GPU-accelerated model serving with Nemotron-3B (64k ctx) for local inference. Gemma 4 for audio. Ollama drop-in for cloud models.

ADR-013

Replica System

Ephemeral agent instances (critic, planner, pipeline roles) with isolated system prompts and named slot routing.

ADR-008

Think-at-Rest

Idle reflection: System 1/2 gap detection, identity and memory consolidation, proactive Telegram thoughts.

ADR-005

User-Request Evolution

Evolution anchored to real user failures. Only fires when a request genuinely fails. Human confirmation gates resolution.

ADR-020

Agent Auto-Discovery

Auto-discovers OpenClaw, coding agents (opencode, Aider, Claude Code), and peers via port scan, mDNS, and MCP.

ADR-015

Perceptive Context

Three-layer composable context: token-aware history, embedding retrieval, and skill-driven context providers built fresh per inference.

ADR-014
Live Status

System telemetry

Online
Status
6.2 VRAM / 16 GB
12d 4h Uptime
35 Skills
3 Peers
2h ago Last Evolve
0 Active Replicas
0 Queue
View Full System Status →
Get Started

Install and run locally.

terminal
# Clone the repository
$ git clone https://github.com/fabiopacifici-bot/kernel-evolving.git
$ cd kernel-evolving && bash install.sh

# Install the EVA CLI
$ ln -sf "$(pwd)/eva" ~/.local/bin/eva

# Start the agent
$ bash start.sh && eva /status

Kernel Evolving runs entirely on your own hardware. Everything is local-first with cloud escalation only for code synthesis. No data leaves your machine unless you explicitly allow it.

Requirements Python 3.11+, CUDA GPU (8GB+ VRAM)

Read the Docs Latest Release

Skills Catalog

35 skills

Every skill is autonomously generated, tested, and curated by the agent. Each one lives in its own git branch with a full SKILL.md spec, command interface, and test suite. Skills are synced from kernel-synthesized-skills .

No skills synced yet.

Run php artisan skills:sync to fetch the catalog from GitHub.