cld

Command line daemon. A fast, persistent assistant for the mindful.

You don't need much to connect your messaging app to a coding agent. Just processes, persistent memory, pipes, and poll. Written in Zig.

Reduce the amount one must remember. — Zig Zen

Single static binary. ~500KB.

Install

brew tap vincehodsdon/tap
brew install cld
brew services start cld

Grant Full Disk Access to the installed binary in System Settings > Privacy & Security.

Architecture

Adapter  ──>  Provider  ──>  LLM
    │                         │
    │        Memory           │
    │     (markdown files)    │
    <────────────────────────<

• Adapter: message transport (iMessage via imsg CLI)
• Provider: LLM backend (Claude via claude CLI, headless multi-session)
• Memory: file-based context — identity, personality, long-term notes, daily logs and summaries

Adapters

Adapter	Transport	Dependency
iMessage	imsg CLI	imsg

Providers

Provider	Backend	Dependency
Claude	claude CLI	Claude Code

Configure

~/.config/cld/config.json

{
  "adapters": {
    "imessage": {
      "allowed_senders": ["+15555555555"]
    }
  }
}

Empty allowed_senders permits all senders.

Status

cld status

Verifies binaries, auth, database access, config, and memory directory.

Memory

Files in ~/.local/share/cld/memory/:

IDENTITY.md           # who the assistant is
SOUL.md               # personality and tone
USER.md               # facts about the user
MEMORY.md             # long-term notes
YYYY-MM-DD.tiny.md    # daily summary
YYYY-MM-DD.full.md    # full conversation log

All optional. Changes picked up automatically via stat checks.

Project Status

Early development. Core loop works — messages come in, get routed to Claude, responses go back out, memory accumulates. Running it daily as a personal assistant over iMessage.

What's solid:

• Event-driven process pool with poll-based I/O
• iMessage adapter (watch + send)
• Claude provider with session reuse and queuing
• File-based memory with stat-checked hot reload

What's next:

• More adapters (Signal, Telegram, etc.)
• More providers (local models, OpenAI, etc.)
• Memory compaction (daily tiny.md generation)
• Skills (checked-in, reviewed prompts the assistant can invoke)
• Tool use and agent capabilities