claude-agent-kit
Interview-driven scaffolder for personal Claude Code agents.
Install on a fresh machine
macOS or Linux. No prereqs — installer auto-fetches git, bun, and Claude Code.
curl -fsSL https://raw.githubusercontent.com/d00ks/claude-agent-kit/main/install.sh | bash
Wizard prompts you for everything: provider (Telegram or Slack), modules, output dir, label. Auto-chains the bootstrap. No flags, no env vars.
What it gives you
- A vault-shaped working directory the agent reads / writes — markdown, inspectable, survives restarts.
- A persona
CLAUDE.mdtuned to an archetype: Executive Assistant or Financial Advisor. - Opt-in capability modules dropped into
.claude/skills/— read at runtime. - Telegram or Slack messaging via the official Anthropic plugins.
- A heartbeat
/loopthat runs on a cron and drains the agent's inbox. - A launchd plist (macOS) so the agent restarts itself.
Capability modules
| Module | Description |
|---|---|
inbox-triage | Urgent / defer / delegate routing. Vacation-mode toggle. |
calendar | Read, create, move, cancel events. Conflict detection. |
scheduling | Multi-party scheduling with time-zone math. |
follow-up-tracking | Lightweight CRM — last-contact + next-action per relationship. |
meeting-prep | Pre-meeting context bundles 30 min before each meeting. |
comms-drafts | Draft replies in the operator's voice. Drafts only (V1). |
light-research | Web / doc fetch + summarize. |
google-workspace | Gmail / Drive / Sheets / Docs / Calendar via the gog CLI. Requires a GCP project. |
google-workspace-mcp | Gmail / Drive / Calendar via Anthropic's hosted MCP. 60s OAuth, zero GCP setup. |
Architecture
A flat, inspectable markdown tree the agent reads + writes. Restart-survivable because state lives on disk.
your-vault/
├── CLAUDE.md # root agent instructions
├── INDEX.md # vault map
├── HANDOFF.md # operator runbook
├── bootstrap.sh # one-shot setup
├── personas/
│ └── <persona-name>/
│ ├── CLAUDE.md # archetype-templated instructions
│ ├── inbox/ # tasks others drop here
│ ├── processing/ # in-flight work
│ └── memory/ # daily logs
├── shared/
│ ├── MEMORY.md # canonical operator preferences
│ └── memory/
│ ├── active-tasks.md
│ ├── lessons.md
│ └── feedback.md
└── .claude/
├── skills/<module>/SKILL.md
└── plugins/<provider>/
How agents adapt over time
Every operator correction + preference is first-class data. The agent treats workflow rule tuning (triage thresholds, escalation criteria, routing logic, delegate rules, sensitive-topic list) as its own responsibility — it edits shared/MEMORY.md in operation, no human-in-the-loop. Memory hygiene is a hard rule: state changes get logged to today's daily note before the user-visible reply.
Onboarding is conversational + iterative. The agent asks the operator about themselves over the first few sessions and fills in {{placeholders}} in its own CLAUDE.md as it learns.