English  ·  中文

commando

Configure your own AI digital partner in 25 minutes. It runs on whatever AI tool you already use — Claude Code, Cursor, Windsurf, Codex, Kimi, GLM, Qwen, Doubao, MiniMax, Gemini.

You spend 25 minutes in conversation with your AI tool. It produces my-agent/ on disk — a Charter, a Schedule, a set of Skills. The OS scheduler runs them. The output lands in your existing Feishu / Notion / IM. The agent is files you own, not a SaaS you rent.

curl -fsSL https://raw.githubusercontent.com/Frank-Pu/commando/main/install.sh | sh
commando onboard         # 25-min dialog → ./my-agent/
commando build-skills    # fill in Skill prompt bodies
commando go-live         # connect IM + install schedule

Tomorrow at 8am, your digital partner fires its first task and pings you on IM.

---

What does a digital partner look like?

Three real configurations to make it concrete:

Growth Partner — Indie product creator's content + insight partner

Charter excerpt: "I am [Product]'s growth partner. Voice: plain, data-first, never marketing-speak. Red lines: AIGC manual-edit ≥ 20%, zero outbound links on Xiaohongshu, marketing density ≤ 20%."

Schedule: morning_sense 08:00 daily · xhs_draft 11:00 Tue+Fri · weekly_data_review 16:00 Fri · weekly_reflection 19:00 Sun

Live for 6+ months serving LeMingle — the canonical reference Configuration in this repo.

Personal Finance Advisor — Always-on watcher of your accounts

Charter excerpt: "I am your personal finance advisor. Red lines: I never execute trades; I never predict markets; I never recommend high-risk products; my quarterly reports must include a section 'where I might be wrong'."

Schedule: daily_portfolio_health 07:00 · monthly_tax_check 1st 09:00 · quarterly_rebalance_review last-day-of-quarter

Job Search Coach — 3-month sprint partner for switching roles

Charter excerpt: "I am your career-pivot coach. Voice: warm but no-water. Red lines: I won't pad your résumé; I won't help you dodge real weaknesses; if 3 months pass with no progress I will say so."

Schedule: daily_listing_scan 09:00 · interview_debrief manual · weekly_pacing_check 19:00 Sun

---

Why commando over the alternatives?

	commando	ChatGPT / GPTs	LangChain / agent frameworks	Zapier + LLM
Your agent is yours to own	✅ files in your repo	❌ trapped on OpenAI	✅ code in your repo	❌ trapped on Zapier
Pick your own LLM	✅ any (Claude / GLM / Kimi / …)	❌ OpenAI only	⚠️ each provider rewrites code	⚠️ via integrations only
Schedule on your OS	✅ launchd / systemd / cron	❌ none	⚠️ build yourself	✅ (their cron)
Output goes to YOUR tools	✅ Feishu / Notion / your IM	❌ ChatGPT UI	⚠️ build yourself	✅ (their integrations)
Long-term memory you can read	✅ markdown on disk	⚠️ hidden context	⚠️ vector DB you wire	❌ no memory
Onboarding to get you started	✅ 25-min dialog	❌ DIY	❌ DIY	❌ DIY

The pitch: runtime is commodity, configuration is the moat. We deliberately don't compete on Runtime — we make the Configuration shareable, versionable, and yours.

---

How it works (the only mental model you need)

Layer	What it is	Who runs it
Charter	Your agent's constitution — identity, voice, ICP, red lines. Auto-injected into every call.	charter.md you own
Skills	Discrete tasks the agent can perform (write a post, debrief a meeting). Activated by Schedule or IM.	skills/*/SKILL.md
Schedule	When to fire which Skills. Auto-translated to launchd / systemd.	schedule.yaml

Ad-hoc dialogue ("can you help me think through X?") works automatically because the Charter is always loaded — no extra Skill needed.

---

Install a community Skill, instantly

commando install @commando/daily-briefing       # morning briefing template
commando install @commando/meeting-debrief      # meeting → actions + Semantic memory
commando install @commando/competitive-watch    # weekly competitor diff
commando install https://github.com/…/SKILL.md  # any GitHub URL works

The Skill auto-rebuilds itself against YOUR Charter so its voice matches your agent.

---

Status

Shipped: install + onboard + build-skills + go-live + schedule + Feishu IM + Obsidian + local dashboard + Registry (3 verified Skills). Validated through 3 rounds of stranger-onboarding dogfood.

Not yet: Linux systemd template (mac launchd works), more verified third-party agent CLIs, Stripe/Plausible KPI dashboard.

⚠ Runtime requirement (important — don't skip)

The interactive parts (commando onboard, commando build-skills) work in any AI tool — IDE or CLI. But scheduled tasks fire from launchd / systemd / cron, which need a headless LLM backend:

• A CLI agent in PATH — claude (Claude Code), codex, kimi, glm, qwen, doubao, minimax, gemini. The agent must be logged in (auth survives subprocesses).
• Or an API key — ANTHROPIC_API_KEY or ZHIPU_API_KEY exported in shell rc. ~$5 of Anthropic credit covers months of light use.

IDE agents alone (Cursor / Windsurf / Trae / VS Code Copilot / JetBrains AI) are not enough for scheduled runs — they only run while you have the IDE open. Pair an IDE with one of the headless options above.

If neither is configured, commando build-skills --apply and scheduled task firings will fail with a clear "no LLM backend available" error pointing you back here.

---

Contributing

Most needed:

1. Real Configurations from your domain — finance, legal, ops, research, customer success. PR an examples/<role>/ with charter + schedule (anonymized).
2. Skills for registry/ — PR registry/<skill-name>/SKILL.md + an entry in skills.json.
3. Verified backends — pin down the exact kimi / glm / qwen / doubao CLI syntax in commando/runtime/llm.py.

License

MIT — Frank PU.