KERNEL
Permanent memory and multi-agent orchestration for Claude Code.
Claude Code forgets everything when you close it. KERNEL gives it memory that persists forever—what worked, what broke, where you left off. Every conversation makes Claude Code smarter about YOUR project.
Install
Claude Code (Terminal)
/plugin marketplace add ariaxhan/kernel-claude
/plugin install kernel
/kernel:init
Claude Desktop
- Open Customize (sidebar) → Personal plugins
- Click + → Add marketplace from GitHub
- Enter:
ariaxhan/kernel-claude
- Click the KERNEL plugin → Install
- In a project, run
/init to set up memory
Cursor
Install via Claude Code or Claude Desktop first. Cursor shares the same plugin configuration automatically.
See docs/QUICKSTART.md for the full setup guide.
What It Does
Memory
Claude Code remembers everything across sessions. What worked. What broke. Where you left off. Lessons learned. It's all saved in _meta/ and persists forever.
Agents
Specialized helpers for different tasks:
- Surgeon - Builds features with minimal changes
- Adversary - QA agent, finds edge cases and bugs
- Reviewer - Code review with >80% confidence threshold
- Researcher - Looks up solutions before building
- Scout - Explores and maps your codebase
- Validator - Pre-commit checks (tests, lint, security)
Tiered Routing
KERNEL automatically routes your requests:
| Tier | Files | Approach |
|---|
| 1 | 1-2 | Execute directly. No subagents needed. |
| 2 | 3-5 | Spawn surgeon agents to implement. |
| 3 | 6+ | Surgeon agents implement; adversary agents verify. |
Skills
Skills are methodologies that load on-demand. When you mention "debug" or "security," the relevant skill loads automatically. Skills inform HOW agents work.
| Skill | What It Does |
|---|
debug | Scientific debugging: reproduce, hypothesize, binary search to root cause |
testing | Test behavior not implementation; edge cases over happy paths |
tdd | Test-driven: red-green-refactor cycle, tests before code |
security | Input validation, auth, secrets management, OWASP top 10 |
api | REST design: resource naming, status codes, pagination, versioning |
backend | Repository pattern, caching, queues, N+1 prevention |
e2e | Playwright: Page Object Model, flaky test strategies |
refactor | Behavior-preserving transforms; tests green before AND after |
design | Frontend aesthetics; break generic AI patterns |
architecture | System design, modules, dependencies, coupling |
git | Atomic commits, conventional messages, branch strategies |
context-mgmt | Token management, compaction strategies, handoffs |
orchestration | Multi-agent coordination, contracts, fault tolerance |
performance | Measure before optimizing; identify real bottlenecks |
eval | Eval-driven development for AI workflows |
Skills live in skills/{name}/SKILL.md. Each has a reference doc in skills/{name}/reference/.
Daily Use
Start every request with /ingest
This is the universal entry point. It reads memory, classifies your task, picks the right approach, and routes to the right agent. Always start here.
/ingest add user authentication to the app
Or just type /ingest and describe what you want on the next line.
Note: In Claude Code terminal, commands use the kernel: prefix (/kernel:ingest). In Claude Desktop and Cursor, they appear without the prefix (/ingest). Same functionality, different naming.
Check: /kernel:validate before committing
Save: /kernel:handoff before closing
Commands
| Terminal | Desktop/Cursor | What It Does |
|---|
/kernel:ingest | /ingest | Guided flow — classify, scope, execute with human confirmation |
/kernel:forge | /forge | Autonomous engine — heat/hammer/quench/anneal until antifragile |
/kernel:dream | /dream | Creative exploration — 3 perspectives, 4-persona stress test |
/kernel:diagnose | /diagnose | Systematic debugging + refactor analysis before fixing |
/kernel:retrospective | /retrospective | Cross-session learning synthesis + pattern promotion |
/kernel:metrics | /metrics | Observability dashboard — sessions, agents, hooks, learnings |
/kernel:validate | /validate | Pre-commit quality gates |
/kernel:tearitapart | /tearitapart | Critical pre-implementation review |
/kernel:review | /review | Code review for PRs |
/kernel:handoff | /handoff | Save progress for next session |
/kernel:init | /init | Setup (run once per project) |
/kernel:help | /help | Show help |
Local Development (For Contributors)
If you're developing KERNEL locally, symlink the cache to avoid stale copies:
ariaxhan