Kernel is a skill and agent publishing tool. You author skills and agents once, in a neutral format, and kernel deploys them to whatever AI tools you use — Claude Code, Codex, GitHub Copilot, Pi.

Most AI tools have their own directory for skills and agents — .claude/skills/, .codex/skills/, etc. Each expects a slightly different format. Without kernel, you'd have to maintain separate copies for each tool and keep them in sync manually.

Kernel gives you one source of truth and handles the rest.

src/templates/skills/kernel-review/  →  discover  →  render per tool  →  ~/.claude/skills/kernel-review/SKILL.md
                                                                        →  ~/.codex/skills/kernel-review/SKILL.md

Skills and agents live in src/templates/skills/ and src/templates/agents/. Each is a directory with a SKILL.md or AGENT.md file — a neutral description of what the skill does, independent of any specific AI tool.

No manifest, no registration. Add a folder with a SKILL.md inside, and kernel sync picks it up.

It checks your home directory for .claude, .codex, .copilot, .pi. If the folder exists, that tool gets the skills.

Each AI tool has an adapter — a small translator that knows the format that tool expects. The Claude adapter writes SKILL.md files with specific frontmatter. The Codex adapter writes .toml files. Copilot gets .agent.md files. Same content, different format.

Output files are written into the tool directories. Kernel also maintains a sync manifest so it can remove files whose templates have been deleted — no orphans left behind.

Each repo can also have a .kernel/ directory — a committed project memory for goals, tasks, and knowledge that lives in the repo itself rather than in chat history.

.kernel/
  work/
    goals/<id>/goal.md
    tasks/active/<id>/task.md
    tasks/archived/<date>-<id>/task.md
  knowledge/
    notes/
    guides/
    learnings/
  state.json

kernel sync                              # deploy all skills/agents to installed tools
kernel doctor                            # check what's installed and what's out of sync
kernel goal new "make onboarding fast"
kernel task new "write setup guide" --goal <id>
kernel task status
kernel task done <item>
kernel knowledge list

npm install -g @hackefeller/kernel
kernel sync
kernel goal new "make onboarding effortless"
kernel task new "write setup guide" --goal make-onboarding-effortless
kernel task status

just ci                 # typecheck, test, build, binary validation
just validate-binary    # smoke test the compiled dist/kernel
just install            # build and install to ~/bin/kernel

just version-dry-run patch
just release-dry-run
just publish-dry-run

just version-bump patch
just release confirm=true
just publish confirm=true

• Define once, deploy everywhere
• No manifests — discovery is filesystem-based
• Host-specific behavior lives in adapters, not templates
• .kernel/ is the repo's committed project memory
• One markdown file per record; frontmatter is metadata

• Skill metadata compatibility matrix
• Supported skill metadata by provider