convez.ai
AgenteAutomaçãoAvançado
bradbushSFAI

Nolty — Agentes Claude sempre ativos sem créditos de SDK

bradbushSFAI/nolty

Agentes cron no seu plano Claude padrão — sem créditos de SDK

Ver no GitHub
InicialPluginScore · de 0 a 100

Framework que torna possível rodar agentes Claude sempre ativos (style OpenClaw) usando exclusivamente Claude Code interativo, sem consumir créditos do Agent SDK. Inclui dispatcher cron de 15 min, integração Telegram, sub-agentes com herança de tarefas e auto-recuperação quando Claude Code atualiza.

Instalação

$npx convez-ai bradbushSFAI/nolty

1 comando no terminal: detecta o formato do repositório (marketplace ou plugin avulso) e instala via Claude Code. Requer Node 18+ e o Claude Code instalado.

Métricas

Estrelas
1
Forks
0
Acompanhando
0
Saúde
57%
Último commit
há 5 meses
Versão
v0.1.0
Issues abertas
0
Contribuidores
1

Casos de uso

  • Migrar de OpenClaw para Claude Code mantendo sempre-conectado
  • Escalonar tarefas recorrentes (resumos diários, rollups semanais, atualizações de planilhas)
  • Monitorar métricas web com cron signado (LinkedIn analytics, dados de BI)
  • Responder a comandos do Telegram em qualquer lugar através de um agente Claude persistente

Veredito

Nolty resolve um problema concreto pós-mudança de política da Anthropic (Jun 2026): rodar agentes sempre-ativos sem drenar créditos de SDK. Arquitetura robusta baseada em padrões reutilizáveis (tmux, cron-runner, sub-agentes, MCP inheritance). Recomendado para quem vinha usando OpenClaw e precisa migrar mantendo ergonomia.

README

A tradução para o português deste README ainda está na fila.

Nolty

Claude Running Like OpenClaw — always-on AI agents on your standard Claude plan

Cron-driven AI agents on your standard Claude plan. No Agent SDK credits required.

Nolty is a migration toolkit for OpenClaw users moving to Claude Code, and a complete always-on agent stack for anyone who wants OC-style ergonomics on top of Claude Code: an always-on Claude session that listens to Telegram, runs scheduled crons every 15 minutes, dispatches work to Task sub-agents (so the main thread stays responsive to chat), and self-heals when Claude Code auto-upgrades.


Why this exists

If you used OpenClaw, you got used to:

  • An always-on agent you could ping from Telegram anywhere
  • Scheduled background jobs (morning brief, daily recap, weekly rollups)
  • The agent owning its own memory and knowing your week
  • Web-driven crons (LinkedIn analytics, sheet updates) via a real signed-in browser

Claude Code can do all of this — but you have to wire it. Out of the box you get the CLI, sub-agents, MCP plugins, claude-in-chrome, and slash commands. What's missing is the always-on session, the cron dispatcher, the dispatch routing rules, and the recovery patterns when things drift.

Nolty packages all of that.


Key headline: no Agent SDK credits

Anthropic's June 15, 2026 policy change separates Claude Agent SDK usage (claude -p / non-interactive) from interactive Claude Code. SDK usage now draws from a separate monthly credit pool; interactive Claude Code stays on your standard plan.

Nolty's architecture is 100% interactive Claude Code. The cron-runner is a Python script that types slash commands into a live tmux session via tmux send-keys — it never calls claude -p. Sub-agents run inside the interactive session. The Telegram plugin runs inside the interactive session.

Anyone running an OpenClaw-style headless cron setup will start drawing from the Agent SDK credit pool after June 15. Anyone running Nolty won't.


What you get

Seven reusable patterns, each independently valuable:

  1. cron-runner/ — Python LaunchAgent that wakes every 15 min, reads cron-jobs.json, and types due slash commands into Nolty's tmux session. Replaces runCLAUDErun. No menu-bar GUI, JSON config, programmable.

  2. TelegramConfig subfolder pattern — Plugin scope is workspace-root only. Putting the Telegram-enabled settings in a subfolder lets you open the parent in an IDE without spawning a duplicate listener.

  3. Cron-suffix dispatch routing — CLAUDE.md teaches the agent that when a slash command arrives with [cron model:X effort:Y], it spawns a Task sub-agent (with that model + effort) rather than running in the main thread. Sub-agents inherit claude-in-chrome MCP — Chrome-driven crons just work.

  4. cron-management skill — Natural-language CRUD: "show crons", "disable heartbeat", "run audible deals now", "add a cron that pulls the weather at 8am". Live management without editing JSON by hand.

  5. Absolute-path discipline — Sub-agents get a thin PATH. Skills use /opt/homebrew/bin/gog (absolute) instead of gog. This single rule prevents the most common silent-failure mode.

  6. /nolty-restart recovery skill — Global slash command that diagnoses and restarts the whole stack from any Claude Code session. When Telegram goes silent or crons stop firing, you run this from any terminal and it figures out what's broken and fixes it.

  7. Heartbeat self-heal — The every-30-min heartbeat skill (STEP 0.5) checks for Claude Code version drift and auto-restarts the tmux session before you notice. Catches the most common "everything went silent after CC auto-upgrade" failure.


Quick start

# 1. Clone
git clone https://github.com/bradbushSFAI/nolty.git ~/Documents/CodingProjects/nolty
cd ~/Documents/CodingProjects/nolty

# 2. Install deps
brew install tmux python   # if not already installed
pip3 install croniter

# 3. Set up Telegram bot (see docs/SETUP_TELEGRAM.md)
#    - Create bot with @BotFather
#    - Get your chat_id from @userinfobot
#    - Save token + chat_id to ~/.claude/channels/telegram/.env (or ask Claude to do it)

# 4. Copy templates to live files (or ask Claude to integrate your OpenClaw
#    or other agent files into these files as they are copied over)
cp USER.template.md USER.md
cp MEMORY.template.md MEMORY.md
cp IDENTITY.template.md IDENTITY.md
cp SOUL.template.md SOUL.md
cp TOOLS.template.md TOOLS.md
cp HEARTBEAT.template.md HEARTBEAT.md
cp AGENTS.template.md AGENTS.md
cp cron-runner/cron-jobs.example.json cron-runner/cron-jobs.json

# 5. Edit USER.md, MEMORY.md, cron-jobs.json with your specifics (or ask
#    Claude to help — paste your context, let her fill in the placeholders)

# 6. Install the LaunchAgent (one-shot install script)
./scripts/install.sh

# 7. Start Nolty's tmux session
./clawd-restart.sh

# 8. Verify everything
./scripts/preflight-check.sh

You should now be able to message your Telegram bot and get a reply. Within 30 minutes the first heartbeat fires.

For the full setup walkthrough including BotFather, LaunchAgent persistence, and claude-in-chrome configuration, see docs/SETUP_TELEGRAM.md and docs/SETUP_LAUNCHAGENT.md.


Repository layout

nolty/
├── README.md                  # this file
├── LICENSE                    # MIT
├── CLAUDE.md                  # Nolty's operating rules (loaded every session)
├── *.template.md              # foundation file templates (copy → *.md)
├── clawd-restart.sh           # restart Nolty's tmux session
├── assets/
│   ├── nolty.png              # Nolty the orange crab mascot
│   └── sfai-logo.png
├── TelegramConfig/            # subfolder with Telegram plugin enabled
│   ├── CLAUDE.md
│   └── .claude/settings.json
├── cron-runner/
│   ├── bin/
│   │   ├── cron-runner.py     # the dispatcher
│   │   └── send-telegram.sh   # emergency fallback only
│   ├── cron-jobs.example.json # sample jobs
│   ├── requirements.txt       # croniter
│   ├── com.example.cron-runner.plist.template
│   ├── state/                 # gitignored runtime state
│   └── logs/                  # gitignored runtime logs
├── skills/                    # core skills (cron-management, heartbeat, nolty-mood, chatgpt-image)
├── examples/                  # 14 example skill patterns (Brad's originals, genericized)
├── .claude/
│   ├── settings.json          # plugin DISABLED at parent scope
│   └── commands/
│       └── nolty-restart.md   # /nolty-restart global recovery command
├── memory/                    # gitignored daily session logs
├── tests/
│   └── test_cron_runner.py    # pytest tests
├── scripts/
│   ├── install.sh
│   ├── preflight-check.sh
│   └── scrub.sh
└── docs/
    ├── MIGRATION_FROM_OPENCLAW.md
    ├── SETUP_TELEGRAM.md
    ├── SETUP_LAUNCHAGENT.md
    ├── SETUP_GOG.md
    ├── SETUP_CLAUDE_IN_CHROME.md
    ├── RECOVERY.md
    ├── SKILL_REFERENCE.md
    ├── CUSTOMIZING_NOLTY.md
    ├── TESTING.md
    ├── PORTING.md
    ├── architecture.md
    └── cron-runner-internals.md

Requirements

Required (the baseline):

  • macOS (Linux port possible — see docs/PORTING.md — Windows requires WSL2)
  • Python 3.9+
  • croniter (pip install croniter)
  • tmux (brew install tmux)
  • Telegram bot (free; create with @BotFather)
  • telegram@claude-plugins-official plugin (installs via claude CLI)
  • gog CLI (Gmail/Calendar/Sheets/Drive) — see docs/SETUP_GOG.md

Optional (per skill):

  • claude-in-chrome configured — required only for skills that drive the browser (LinkedIn examples). Free; install the Chrome extension, sign into your accounts.
  • OpenAI API key + uv — required only for image-generation skills (chatgpt-image, used by nolty-mood). See docs/SETUP_OPENAI.md.
  • An Obsidian-style notes folder — many examples reference YOUR_OBSIDIAN_VAULT (meeting prep, daily-recap, linkedin-monthly-analysis, content-hour). Any directory of markdown files works — see docs/SETUP_OBSIDIAN.md.
  • qmd — local markdown vector search; only the qmd-reindex example needs it
  • audible-deals Python package — only the audible-deals example needs it
  • Perplexity MCP — only the home-loan-rate example needs it
  • Zapier MCP — only needed if you adapt daily-recap to pull Zoom transcripts via zoom_get_meeting_summary
  • A CRM — meeting-prep example has an optional step for Folk/HubSpot/Apollo/etc. lookup; skip if you don't use one

Not a dependency: claude-in-chrome is built into Claude Code, not a separate install. OpenClaw is not a dependency (this toolkit replaces it).


What about Linux / Windows?

macOS is the only OS supported in v1. The macOS-specific bits are:

  • launchctl + LaunchAgent plist (replace with systemd user units on Linux)
  • ~/Library/LaunchAgents/ boot persistence (replace with systemctl --user enable)
  • A few lsof / pgrep calls (work on Linux; need WSL on Windows)

A Linux port is ~1 day of work for a contributor. See docs/PORTING.md for the substitution table.


Documentation


Status

Pre-release. Public v0.1.0. Tested on macOS 24.6.0 (Darwin), Python 3.11, tmux 3.4. Battle-tested in a single-user setup since May 2026.


Contributing

Issues and PRs welcome. Focus areas:

  • Linux port — systemd user unit equivalent of the LaunchAgent
  • Alt channels — iMessage, Discord, Slack plugin support
  • More example skills — yours might help someone else

See CONTRIBUTING.md for the standard PR workflow.


License

MIT — see LICENSE. Copyright 2026 Brad Bush and Strategy For AI.


Built by Strategy For AI

Strategy For AI helps professional services firms — CPAs, lawyers, engineers, consultants, PE/investment banking, and architecture firms — adopt AI in ways that actually move the business. We focus on friction removal: identifying the workflows that drain hours each week, designing AI tooling that fits how your team actually works, and shipping it in weeks not quarters.

We built Nolty because we needed an always-on agent that handled our own scheduling, content, and operations — and we wanted to share what worked.

If your firm wants to put AI to work in the same disciplined way, strategyfor.ai is where to start.

Ver o original no GitHub

Bom para

Relacionados