Cully
A self-improving harness for coding agents and software projects.
Website · Documentation
Cully runs Claude Code, Codex, Cursor and other coding agents inside one terminal runtime, adding session health, advisor guidance, private project memory and shared team workflows. Each agent keeps its own reasoning and tools; Cully supplies the session around it. It also scores itself every five hours and proposes fixes, which it writes only after you approve them.
Why Cully
Useful context should survive a session. Cully surfaces repeated failures and unchecked edits, and brings back saved decisions and next steps when you switch sessions or agents. See what happened, then continue with the context that matters.
Get started
Install on macOS or Linux:
curl -fsSL https://cully.net/install.sh | sh -s -- --agent codex
Open a new terminal, then:
cully setup --all # start local memory services and connect your agents
cully run codex # or claude, cursor, another executable
Setup is user-wide: run it once, then use Cully in any project. No project .cully folder is required. The local memory stack requires Docker. For an existing server, use cully setup --mcp-url URL. See the quickstart for prerequisites and other agents.
Cully supports native terminals on macOS and Linux with zsh + Oh My Zsh. Setup installs the required shell and Nerd Font preset automatically, with backups. Use macOS Terminal, Ghostty, iTerm2, Kitty, WezTerm, Alacritty, Warp or a supported GNOME terminal on Linux. See terminal setup.
What you get
| Capability | What it helps with |
|---|
| Live guidance | Notice context pressure, repeated failures and edits awaiting checks. |
| Terminal compatibility | Install zsh/Oh My Zsh and a Nerd Font preset; verify icons, colors, wheel, divider and advisor behavior. |
| Technique discovery | Hourly, bounded public research on current agent releases, workflows, skills and hooks, filtered against your task. Opt out with CULLY_DISCOVERY=off. |
| Harness audit | Find setup gaps in instructions, permissions, memory, skills and hooks. |
| Token impact | Inspect lookup overhead, oversized output and observational savings estimates. |
| Replay and handoff | Review recorded steps, recover a stuck session and prepare the next agent. |
| Project memory | Recall saved decisions, attempts, results and next steps through MCP and Mem0. |
| Self-improvement | Cully evaluates itself on a schedule, proposes fixes from failing checks, and applies one only when you approve it in cully improve. |
| Evals | Score setup, harness, the improvement loop, task efficiency (verified, time, steps), advisor and memory against planted ground truth. |
| Workflow profile |
cully status # session health and pending checks
cully harness # read-only setup audit and recent output baseline
cully replay # recorded session activity
cully handoff --print # a continuation summary
cully improve # review what Cully suggests for itself
cully eval tasks # task efficiency from your own sessions
cully impact --explain # how the net token figure is built
cully update --check # check for a newer CLI
Signals depend on the agent's integrations. Token figures are estimates, not billing totals or proof of savings. See current capabilities.
Solo and Team
- Solo: a coding-agent harness runtime with session guidance, replay and private owner-scoped memory.
- Team: a project management harness runtime for people and AI agents, with shared tasks, checkpoints, independent review and explicitly published lessons. Requires OAuth and project membership; agents launch manually.
Both use the same CLI. Private notes and local journals stay separate from shared project records.
Privacy
The local journal records session metadata, not prompts, file contents or tool output. Historical measurement tools read local Claude transcripts without saving raw transcript text. Set CULLY_JOURNAL_PATHS=0 to omit paths and command names from the journal. Read the privacy guide for storage and model boundaries.
Mobile preview
Cully Self brings dated personal work briefs, local tasks and read-only questions to Android and iOS. Source builds are available; store distribution and device verification remain incomplete.
Learn more
Architecture · Self-hosting · Development · Product direction
Licensed under Apache 2.0.
