claude-code-hermit

claude-code-hermit

Claude Code plugin that turns a Claude Code instance into a 24/7 AI assistant. Orchestrates, Self-improves, Pro-Active, Cost-aware, Observable, One Claude subscription, multiple hermits.

A hermit is a Claude Code session that doesn't end: It wires the native primitives (/loop, CronCreate, Channels, Monitor, Auto-memory, native Tasks, Sub Agents, etc) into an always-on agent that orchestrates, keeps its own state, routines, knowledge and reaches you on Discord, Telegram, or the Claude app when something needs a decision. It can run on your existing subscription, and one subscription runs several hermits.

# Install
claude plugin marketplace add gtapps/claude-code-hermit
claude plugin install claude-code-hermit@claude-code-hermit --scope local

# Boot Claude Code and run the setup wizard
/claude-code-hermit:hatch

# Go always-on
/claude-code-hermit:docker-setup

What you get

Just Claude Code and everything is yours to shape: channels (Discord/Telegram), MCP servers, routines, watches, the heartbeat checklist. Hermit adds persistence, a learning loop, and a quick setup to wire everything.

• Orchestrator instructed to delegate tasks & exploration to other agents, main context stays clean for token efficiency.
• Heartbeat gates /loop behind a filesystem-only precheck so it stops paying the model every tick, sweeping your checklist for zero tokens.
• Routines wrap CronCreate jobs that expire in 7 days and fire in the machine's timezone: they self-rearm daily, run on your wall clock, and are managed by /hermit-routines.
• /watch wraps Monitor streams that die with the session: it auto-starts from config (or plain language) and routes findings to your notifications, silent when quiet.
• Channels let you DM a session; the hermit agent acts on it ("accept PROP-014", "status") and pings you first when something needs a yes/no.
• Auto-memory + knowledge Two layers. Claude Code's native auto-memory holds operator facts and preferences (how to work with you); on top, the hermit adds a raw/ → compiled/ knowledge base — domain outputs and living topic pages updated in place — re-injected as a catalog within a context budget at session start. Your Discord/Telegram DM text is also captured locally, so decisions made over chat outlive the thread: weekly-review distills them into memory (opt out with knowledge.channel_log_enabled: false). /recall searches across all of it.
• Task snapshots persist native Tasks past session end, so the plan survives archives.
• Profile-gated guardrails scope deny patterns + sandbox per profile, locking the unattended agent down harder than the one you're watching.

Sessions self-manage. Daemons auto-archive at 12h idle and at midnight when you're away, so evidence reaches the learning loop without a manual close. An external watchdog restarts dead sessions, nudges wedged ones, re-arms missed schedules, clears stale context after a midnight close, and compacts long-running context so cold wakes don't re-pay the full accumulated history — recovery never depends on the session being conscious.

It reaches you first. Notifications default to a native push (headless-friendly), or a Discord/Telegram DM you can reply to if you've paired a channel.

Cost scales with events, not time. Nothing wakes the model until something happens, so an idle hermit is effectively free.

Learning Loop

A hermit watches what keeps going wrong across sessions, proposes a fix, and asks you yes or no. It won't propose the same thing twice.

At natural pauses — session end, idle ticks, scheduled cadence — it reflects. Most reflections never reach the model: a precheck script gates whether any phase (compute, resolution check, cost spike, digest, newborn) is actually due. When one is, two subagents vet the candidate before it reaches you: