Cue

Behavioral context daemon for AI agents. Watches what you do and publishes structured context so agents start every session informed, not blind.

Quick start

make deps     # fetch dependencies
make build    # compile
make install  # copy to ~/.local/bin/
cue install   # wire up agent hooks

Usage

cue daemon     # Start the background daemon
cue context    # Print current briefing (JSON)
cue health     # Check daemon is running
cue install    # Set up agent integrations

Architecture

┌─────────────────────────────────────┐
│  cue daemon                          │
│                                      │
│  ┌──────────┐  ┌──────────────────┐  │
│  │ watchers  │  │  context builder  │  │
│  │  • git    │──│ → ~/.cue/        │  │
│  │  • procfs │  │   context.json    │  │
│  │  • inotify│  └──────────────────┘  │
│  │  • history│                       │
│  └──────────┘  ┌──────────────────┐  │
│                │  sqlite store     │  │
│                │  → session log    │  │
│                │  → event log      │  │
│                └──────────────────┘  │
│                                      │
│  unix socket (/context, /health)    │
└─────────────────────────────────────┘
         │
         ▼
  ~/.cue/context.json ← any agent reads this

Any agent (Cursor, Claude Code, Codex, Hermes) that reads a file gets Cue context for free.

Fidelity levels

Level	What you get	Token cost
wiretap	Live state (cwd, git, procs)	None
memory	+ local embeddings over past sessions	CPU only
coach	+ LLM summary passes, pattern detection	Uses LLM, net saves tokens

Set in ~/.cue/config.json: {"fidelity": "memory"}

License

MIT