Cully

Your agent's copilot. Your team's workspace.

Website · Documentation

Cully has two modes around Claude Code, Codex, Cursor and the agents you already use. Solo is a copilot for your coding agent and work environment: session health, advice, replay, handoff and private memory. Team connects project tasks, ownership, portable checkpoints, review evidence and explicitly shared lessons across people and agents.

Start with Solo mode or Team mode. Both use the same installed CLI; team operations require OAuth and explicit project access. No new mode flag is needed, and solo notes stay private.

Prepare  →  Work  →  Hand off  →  Review  →  Reuse

Works across agents. Learns across sessions.

Why Cully

A coding agent is good at the task in front of it. It does not see the whole session: that the same test has failed three times, that the edits were never checked, that you solved this last week, or what the previous agent left half done. Cully keeps track of that wider picture.

What Cully does

cully run claude      # or codex, cursor, or any agent on your PATH

Capability	What you get
One terminal for every agent	Claude Code, Codex, Cursor CLI and any executable run in the same wrapper with a live health bar and advisor panel.
Loop detection	"Cully noticed something: the same command failed 3 times after edits."
Advisor	A second pair of eyes: context pressure, missing verification, broad searching, workflow gaps. Click it to open.
Replay and timeline	See what the agent did and how: which files it read, edited, created and deleted, which checks failed, where it looped.
Handoff	A structured summary so a different agent can continue without rebuilding context.
Rescue	Evidence and recovery steps when a session is stuck.
Project memory	Decisions, attempts, checks and next steps saved through MCP, found again by meaning with Mem0.
Team workspace	Project access, a task board and inbox, atomic claims and portable checkpoints through CLI or MCP.
Review and shared learning	Acceptance evidence tied to a revision, independent human approval, private lesson drafts and explicit project publication.

── Codex │ oauth-service:main │ 27m │ Context 63% │ ⚠ Loop 3x │ ● 2 unchecked ── Ctrl+] advisor ──

Every value Cully shows is measured. A value with no data is left out, never guessed. See current capabilities for exactly what ships today and what is only planned.

Team mode

Team workflows require an OAuth-enabled server and explicit project membership. Start agents manually, checkpoint selected context, submit reported evidence and ask a project maintainer who did not contribute to approve it. Lessons stay private until their author publishes them; playbooks require a maintainer and a completed source task. See team workflows for actions and examples. Managed execution and automatic learning remain planned.

Supported agents

Agent	In the Cully terminal	Notes
Claude Code	cully run claude	Context pressure through a hook-fed snapshot
Codex	cully run codex	Native footer instruments: model, context, tokens, quota
Cursor	cully run cursor	Wraps the cursor-agent CLI. The editor keeps hooks and MCP tools.
Any other agent	cully run NAME	Health bar and Git state; tool events if the agent can run Cully's hook

macOS and Linux.

Get started

curl -fsSL https://cully.net/install.sh | sh -s -- --agent codex

Open a new terminal, then run cully setup --all to start the local memory stack and advisor and connect detected coding agents. Use --agent codex to select an agent. Setup logs each stage and reports completion only after all components are ready. Then:

cully run codex        # work as usual, with the health bar and advisor
cully status           # measured session health and risk
cully replay           # what the agent did, and how
cully handoff          # hand the session to another agent
cully rescue           # when a session is stuck

See the quickstart for prerequisites and other agents.

How it works

Developer ─► Cully terminal ─► your agent
                  │
         tool events from hooks
                  ▼
          Session journal ─► Session intelligence ─► Advisor · Replay · Handoff · Rescue
                                      │
                                      ▼
                           Project memory (MCP · PostgreSQL · Mem0)

Your agent talks to its model and tools directly. Cully draws around it and records small events.

Component	What it does
cully	Terminal, advisor, agent setup, session status, replay, handoff and rescue
cully-mcp	Shared memory tools; private single-owner mode by default, optional OAuth
cully-data	Private PostgreSQL API and Mem0 projection worker

See the architecture and how Cully works.

Privacy

Your conversations are not Cully's database. The local session journal keeps the time, agent, kind of tool, pass or fail, project-relative file paths with the operation, and the program and recognized subcommand of a command. It never keeps prompts, file contents, command arguments or tool output. Set CULLY_JOURNAL_PATHS=0 to stop recording paths and commands. See privacy.

Documentation

Quickstart · The Cully terminal · Session intelligence · Advisor · Memory · Self-hosting · Team deployment

Roadmap

Shipped work and proposals are kept apart in product direction: guardrails, a published benchmark, a lighter local install, and learned team workflow.

Contributing

See development. Companies can also deploy Cully with MCP Runtime.

Licensed under Apache 2.0.