Cut tokens. Kill context bloat. Run 20× faster.

Five universal primitives

The framework is five plugins.
Five killer commands.

Each one collapses a specific dance the agent does over and over. Mined from real Claude Code session transcripts, not guessed at.

+ your own service wrappers. 10 in the marketplace ↓

It learns from how you actually work.

Three commands you typed 30+ times this week — ready to scaffold.

skill-plus scan reads your session transcripts and surfaces what's worth a real skill. No guessing.

[apr 22 10:32]  railway logs --service api
[apr 22 10:48]  railway logs --service worker
[apr 23 11:15]  railway logs --service api
[apr 23 14:08]  railway logs --service api
[apr 24 14:31]  psql -h staging
[apr 25 15:12]  railway logs --service worker
[apr 25 15:48]  gh pr view --json
[apr 26 16:02]  psql -h staging
[apr 28 16:55]  gh pr view --json
[apr 30 13:38]  railway logs --service api
... noise. patterns hidden in 1,400 lines.

$ skill-plus scan
✓ 3 patterns mined

# ready to scaffold →

Then it turns those patterns into deterministic scripts.

Three files in your repo. One of them runs.

Stdlib Python, pre-wired argparse, envelope contract, secret redaction. You fill in the killer-command body.

{
  "key":        "railway logs --service",
  "count":      14,
  "sessions":   3,
  "first_seen": "2026-04-22",
  "last_seen":  "2026-05-01",
  "args_seen": [
    "--service api",
    "--service worker",
    "--service migrate"
  ]
}

# useless on its own.
# just data about a habit.

├── SKILL.md
├── bin/railway-logs      ← stdlib python
└── README.md

────  bin/railway-logs (excerpt)  ────

def cmd_logs(args):
    out = railway.logs(svc=args.service)
    if args.errors:
        out = [l for l in out
               if l.level == "error"]
    print(json.dumps(out))

# real script. no LLM loop.
# every future session uses it.

Then you use it. Every session.

One clean call where 14 used to be.

Same skill, every service, forever. No token cost beyond the call.

$ railway list | grep api
$ railway service api
$ railway logs

[1,402 lines, every level,
 no time filter, no service
 grouping...]

$ # need errors, last hour
$ railway logs --tail 1000 \
    | grep -iE "err|exception" \
    | head -50

[still messy, no timestamps,
 repeat for worker + migrate]

$ railway-logs --service api --errors --since 1h

✓ 3 errors in api (last 1h)

  10:42  ECONNREFUSED at db.connect()
  10:48  TimeoutError at /api/users/:id
  11:03  429 RateLimit from openai

# one call. same skill, every service.

The marketplace convention

Anyone can publish their own.

<user>/agent-plus-skills on GitHub. Borrowed from Homebrew taps and the GitHub Actions marketplace. No central registry to run.

                  ┌─────────────────────────────────────┐
                  │   osouthgate/agent-plus  (this repo)
                  │   The 5 universal primitives
                  └────────────────┬────────────────────┘
                                   │
                  agent-plus-meta marketplace install
                                   │
                                   ▼
        ┌────────────────────────────────────────────────────┐
        │  <user>/agent-plus-skills  (anyone can publish)
        │  github-remote, vercel-remote, supabase-remote,
        │  railway-ops, linear-remote, openrouter-remote,
        │  langfuse-remote, hermes-remote, coolify-remote,
        │  hcloud-remote  (the reference marketplace)
        │  + your own at <your-handle>/agent-plus-skills
        └────────────────────────────────────────────────────┘

Reference marketplace

10 service wrappers, ready to install.

Talk to your stack from Claude Code. Coolify, Railway, Vercel, Supabase, Linear, GitHub, and more.