Give your AI coding agent a memory. Teach it your project's patterns as you build.
AI agents start every session knowing nothing about your project. Agent Evolver gives them skills that learn from your work — recording decisions, extracting patterns, and evolving over time.
Quick Start
# Run without installing npx @adhd8k/agent-evolver install # Or install globally npm install -g @adhd8k/agent-evolver agent-evolver install
That's it. Agent Evolver detects your coding agent (Claude Code, Cursor, Aider) and installs the skills locally into your project.
What Gets Installed
Five skills your agent can use automatically:
| Skill | What it does |
|---|---|
consult-knowledge |
Looks up past decisions and patterns before acting |
record-architectural-decision |
Captures why you made a decision, not just what you decided |
extract-pattern |
Recognizes when something has been done 3+ times and codifies it |
update-skill |
Lets the agent improve its own instructions based on what it learns |
autonomy-preferences |
Learns when to ask you vs. decide autonomously |
Skills are installed locally to your project (.claude/skills/ for Claude Code) and tracked in git — so your whole team benefits.
Autonomy Preferences
Tired of your agent asking permission for everything? Or worried it's making decisions without you?
The autonomy-preferences skill lets you configure how autonomous your agent should be. On first use, it asks:
How autonomous should I be when making decisions?
- Ask me first — I want to approve architectural, design, and technical decisions
- Suggest then proceed — Show me what you're thinking, but don't wait for approval
- Fully autonomous — Make reasonable decisions, I'll correct you if needed
- Let me customize — Set different levels per category
Decision Categories
| Category | Examples |
|---|---|
| Architectural | Database schema, service boundaries, API architecture |
| Design | API design, type definitions, module organization |
| Technical | Framework selection, algorithms, performance trade-offs |
| UI/UX | Layout, styling, interactions, accessibility |
| Process | Branching strategy, commit conventions, release decisions |
Modes
| Mode | Behavior |
|---|---|
ask |
Agent asks for approval before proceeding |
suggest-then-proceed |
Agent states its decision and continues without waiting |
autonomous |
Agent decides without asking; you correct if needed |
Preferences are stored in .agent-evolver/autonomy-preferences.yaml and persist across sessions.
CLI Commands
# Install all skills (auto-detects your agent) agent-evolver install # Install a specific skill agent-evolver install -s record-architectural-decision # Install globally to your agent's home directory agent-evolver install --global # Force a specific agent agent-evolver install -a claude-code # See what's available agent-evolver list # Check which agents are detected agent-evolver detect
Local vs Global
Local (default, recommended)
Skills live in ./.claude/skills/ — version controlled, project-specific, shareable with your team.
Global (--global)
Skills live in ~/.claude/skills/ — available everywhere, but not tied to any project.
Supported Agents
- Claude Code (claude)
- Cursor (planned)
- Aider (planned)
Contributing
git clone https://github.com/adhd8k/agent-evolver.git cd agent-evolver npm install npm run build npm test
PRs welcome. See open issues.
MIT License