# agent-plus > Drop-in plugins for Claude Code that turn 127-tool-call dances into 1-tool-call answers. Five framework primitives plus a marketplace convention. Stdlib Python, no SDK, no config file, no auth dance. Cuts tokens, kills context bloat, runs 20× faster than the agent's default cold-start dance. ## Install - [One-line install](https://youragentplus.xyz/install.sh): `curl -fsSL youragentplus.xyz/install.sh | sh` — installs all 5 primitives and chains into the wizard - [GitHub source](https://github.com/osouthgate/agent-plus): MIT licensed, stdlib Python 3.9+, runs standalone outside Claude Code - [Manual install via Claude Code](https://github.com/osouthgate/agent-plus#manual-install): `claude plugin marketplace add osouthgate/agent-plus && claude plugin install @agent-plus` ## Five framework primitives - [agent-plus-meta](https://github.com/osouthgate/agent-plus/tree/main/agent-plus-meta): Workspace bootstrap, env-var readiness, identity cache, marketplace lifecycle. Commands: `init`, `envcheck`, `refresh`, `marketplace install|search|prefer`. The wizard that knows which machine you're on. - [repo-analyze](https://github.com/osouthgate/agent-plus/tree/main/repo-analyze): Collapses the ~67 grep + ~60 ls cold-start dance for unfamiliar repos into one JSON envelope. Returns languages, frameworks, build tools, deps, entrypoints, README structure. 127 tool calls → 1. - [diff-summary](https://github.com/osouthgate/agent-plus/tree/main/diff-summary): One-call PR triage. Returns role + risk per file, public-API change detection, test-coverage gaps. Replaces 5–20 manual Read calls per PR review. - [skill-feedback](https://github.com/osouthgate/agent-plus/tree/main/skill-feedback): Agent self-rates skills after each use. JSONL on disk. `report` aggregates, `submit` files an upstream issue with evidence. Closes the feedback loop on whether skills are working. - [skill-plus](https://github.com/osouthgate/agent-plus/tree/main/skill-plus): Mines the Claude Code session log for repeated commands, scaffolds a real stdlib-Python skill, audits it, promotes it to your marketplace. Commands: `scan`, `propose`, `scaffold`, `inquire`, `promote`. ## Example: what a scaffolded skill actually looks like A scaffolded skill is three files in your repo, one of which runs: ``` .claude/skills/railway-logs/ ├── SKILL.md # how the agent discovers the skill ├── bin/railway-logs # stdlib python, runs deterministically └── README.md # auto-generated docs ``` The script body is real stdlib Python. Pre-wired argparse, envelope contract, secret redaction, layered env resolver — about 200 lines you don't write. You fill in the killer-command body. Real example from the `railway-ops` skill in the reference marketplace: ```python # bin/railway-logs (excerpt) # Stdlib only. Agent invokes the binary directly. No LLM in the loop. def parse_log_entries(raw: str) -> list[dict[str, Any]]: """Parse `railway logs --json` output. Soft-fails on malformed lines.""" entries: list[dict[str, Any]] = [] for raw_line in raw.splitlines(): line = raw_line.strip() if not line: continue try: obj = json.loads(line) entries.append(obj if isinstance(obj, dict) else {"message": str(obj), "_non_object": True}) except json.JSONDecodeError: entries.append({"message": line, "_non_json": True}) return entries ``` Deterministic shape in, deterministic shape out. Same call, same JSON envelope, every session. The agent stops re-discovering what it already discovered last session. ## Marketplace convention - [osouthgate/agent-plus-skills](https://github.com/osouthgate/agent-plus-skills): Reference marketplace with 10 service wrappers — github-remote, vercel-remote, supabase-remote, railway-ops, linear-remote, openrouter-remote, langfuse-remote, hermes-remote, coolify-remote, hcloud-remote - [Convention](https://github.com/osouthgate/agent-plus#the-marketplace-convention): `/agent-plus-skills` — anyone can publish their own collection at their GitHub handle. No central registry. Borrowed from Homebrew taps. - [Trust model](https://github.com/osouthgate/agent-plus#the-marketplace-convention): Install pins the commit SHA. Nothing in the cloned repo runs at install time. First-run review per install. Updates opt-in only. Checksum verification when declared. ## Architecture facts - Stdlib Python only — zero pip dependencies - Each plugin is one binary in `bin/` plus a `SKILL.md` discovery file - Every binary takes `--non-interactive --auto` for deterministic JSON envelope output (frozen public contract) - Agent invokes the binary directly — no LLM in the loop, no token cost beyond the call - Works for Claude Code only (claude.ai web Skills and Cowork are out of scope: no Bash, no filesystem, no plugin loader) ## Voice / philosophy The framework is opinionated. Five principles drive every plugin in it: - **Stdlib Python only.** No third-party deps. If a feature needs `pip install`, it doesn't ship. - **Deterministic.** Same inputs, same JSON envelope, every time. No LLM in the runtime path. - **Frozen public contracts.** The `--non-interactive --auto` envelope shape doesn't change between minor versions. Agent prompts written once stay correct. - **Jurisdictional ownership.** Each plugin owns its slice of the agent's surface; collisions are detected and resolved at install time. - **Discipline as the aesthetic.** The framework refuses to do magic on the user's behalf. It makes the deterministic path fast. ## Status Pre-1.0. Four primitives (`agent-plus-meta`, `repo-analyze`, `diff-summary`, `skill-feedback`) plus the marketplace lifecycle have been dogfooded for months. `skill-plus` is newer (0.4.0). Cold-start orientation and diff triage are production-stable. ## Optional - [90-second tour](https://github.com/osouthgate/agent-plus#a-90-second-tour): Walk-through of all five primitives in real terminal output - [Before/after table](https://github.com/osouthgate/agent-plus#before--after): Concrete comparisons of agent-plus vs the default Claude Code dance - [CONTRIBUTING.md](https://github.com/osouthgate/agent-plus/blob/main/CONTRIBUTING.md): Architecture, conventions, the seven patterns, the envelope contract