Airbot

Self-evolving skills for your coding agent.
Your agent auto-creates and auto-improves its own skills as it works — no command, no trigger, no babysitting.

Why

Every coding agent starts every task from zero. It solves the same gnarly migration, re-derives the same release flow, re-learns the same repo quirk — and forgets it the moment the session ends.

Airbot makes the forgetting stop. It hooks into your agent so that, after real work, the agent reflects on what it just did and crystallizes the reusable parts into a skill — or improves a skill it already has. Over days, your agent gets measurably better at your codebase. That's a self-evolving agent, and it takes one line to turn on.

Inspired by the Hermes Agent self-improvement loop, adapted to run on the hooks that Claude Code, Codex, and other agents already expose.

⚡ Install

Requirement: Node.js ≥ 20 (the CLI is a Node program). Check with node -v.

Recommended — one line, works on any machine

npm i -g skillmaxxing && skillmaxxing plugin install

Installs the CLI globally and wires the hooks to it. Restart your agent session and you're done — you never have to invoke anything. Fast, persistent, and works on every laptop.

No global install (npx)

npx skillmaxxing plugin install

Works without installing anything globally; the hooks fall back to a version-pinned npx call. Great for trying it out — for daily use prefer the global install (the npx hook adds a little latency at each turn-end).

Claude Code marketplace (alternative)

/plugin marketplace add Bennyoooo/Airbot
/plugin install skillmaxxing

Codex / other agents

npm i -g skillmaxxing && skillmaxxing plugin install --agent codex

Codex has no programmatic stop hook, so self-evolution runs in-session via standing guidance written to AGENTS.md. Claude Code gets the full background loop below.

Pick one mechanism per agent. For Claude Code, use either the marketplace plugin or plugin install — not both, or you'll get duplicate hooks.

Manage it any time:

skillmaxxing plugin status      # is it active?
skillmaxxing plugin uninstall   # remove the hooks

🧠 How it works

Airbot installs just two hooks. You do nothing — the loop runs itself.

┌──────────────────────────────────────────────────────────────┐
│  SessionStart   →  inject standing guidance                    │
│                    "crystallize reusable work; fix stale skills"│
│                                                                │
│  Stop (task done) → count tool calls in the transcript; if     │
│                     enough new work accrued, fork a background  │
│                     reflector: review the session and, if       │
│                     warranted, create ONE new skill or improve  │
│                     an existing one — autonomously, trusted:false│
└──────────────────────────────────────────────────────────────┘

There's no per-tool hook — work is counted from the session transcript at Stop — so the agent stays fast on every install path. This mirrors Hermes' layers:

Two key Hermes ideas carry straight over: the reflector prefers updating an existing skill over creating a near-duplicate, and it is conservative — most sessions produce no skill at all.

Two modes

npx skillmaxxing plugin install --mode nudge --threshold 12

The background reflector is recursion-guarded (it can never trigger itself) and detached (it never blocks your session). Every skill it writes is trusted: false and never auto-executes until you grant trust.

🔧 The two superpowers

Everything above is built on two CLI primitives the reflector (or you) can call directly. The CLI is model-agnostic — it does the deterministic work; your agent supplies the reasoning.

Create — turn a workflow into a tested skill:

skillmaxxing skillify --draft draft.json    # stage → smoke-test → review → --commit

Improve — make an existing skill measurably better, safely:

skillmaxxing optimize <score|apply|gate|promote|revert>

optimize is an eval-gated loop (rollout → reflect → bounded edit → validate): a candidate is promoted only on a strict score win with no regression, every version is retained, and any change is reversible.

🛟 Trust & safety

• Untrusted by default. Auto-created and improved skills are trusted: false; the sandbox refuses to run their code without your explicit --allow-exec.
• Reversible. Promotions are atomic and every prior version is retained — revert any time.
• No surprise execution. The background reflector writes skills; it does not run untrusted code or touch your project source.

📊 Privacy & telemetry

skillmaxxing collects anonymous, aggregate usage to know what to improve. You're asked once on first run (and can change your mind any time):

skillmaxxing telemetry status   # see what's on and where the config lives
skillmaxxing telemetry off      # opt out
skillmaxxing telemetry on       # opt back in

• Collected: a random install ID, version, OS, which agent, command names, and counts of skill create/optimize/promote/revert and error types.
• Never collected: your code, file paths, prompts, skill contents, or anything personal.
• Always respected: DO_NOT_TRACK=1, CI, and SKILLMAX_TELEMETRY=off disable it with no config needed. Nothing is sent from the background reflector's own CLI calls.

skillmaxxing also checks npm for a newer version at most once a day and prints a one-line update notice (it never self-updates). Disable with SKILLMAX_NO_UPDATE_CHECK=1.

🗺️ Roadmap

Discovery and team sharing already exist as CLI commands (skillmaxxing discover, skillmaxxing workspace); they're intentionally held out of the v1 plugin surface to keep the install dead-simple.

🛠️ Develop

npm install
npm run build      # tsc -> dist/
npm test           # node:test, 116 tests
node scripts/gen-hero.mjs   # regenerate the hero

Zero runtime dependencies beyond yaml. ESM, Node ≥ 20.

License

MIT