ContextOS / Checkpoint

Agents don't hand context to each other.

They read and write through ContextOS. Your context pack follows the work across every tool switch — no recap, no reconstruction.

Continuity layer

Claude → ContextOS → Codex → ContextOS → Claude Code → ContextOS → Antigravity → ContextOS → Claude

Not another coding agent. The continuity layer beneath them.

ContextOS keeps durable project memory where the repo can own it: readable files, explicit tasks, handoffs, decisions, and projections that existing tools already read.

1. Write

Local project memory

Keep constraints, decisions, active work, and handoffs in .contextos/ so the next session starts from something concrete.

2. Inspect

Files before prompts

Open the Markdown, read the generated pack, and choose what moves into the next AI tool.

3. Continue

Agent-shaped output

Ask for a pack aimed at Codex, Claude, Claude Code, Cursor, or a generic agent.

4. Resume

Less recap, more work

The next session gets the task, files to inspect, current status, and the next action without a long spoken reset.

A continuation pack is not a pitch deck.

It is a compact, readable handoff: what changed, what matters, where to look, and what the next agent should do first.

# Continuation Pack

Current task: TASK-004 improve empty-project recovery
Previous agent: codex
Next agent: claude-code

## Files to inspect
- .contextos/tasks/active/TASK-004.md
- .contextos/handoffs/latest.md
- .contextos/context/constraints.md

## Next action
Reproduce the empty-project path and improve the recovery message.

Local trust before automation.

The first workflow is intentionally small: no account, no hidden database, no opaque project state. The repo gets readable context files, and you decide what to share with another tool.

.contextos/
  context/
  tasks/
  handoffs/
  decisions/
  state/

Designed for inspection

The generated output uses obviously readable Markdown. The demo and examples use fake project names and safe placeholder values.

Every new session starts from scratch. Until now.

AI coding tools don't share context with each other. When you switch from Claude to Codex to Claude Code, you spend the first minutes of every session re-explaining what you were doing — burning tokens on setup instead of work.

Problem

Context window wasted on recap

A cold-start prompt that re-explains your architecture, active task, and recent decisions can consume thousands of tokens before a single line of code is written.

Problem

State lost between tool switches

Plan in Claude, hand to Codex, debug in Claude Code — each switch drops the shared memory. You carry it in your head, re-typing it from scratch each time.

With ContextOS

Fewer tokens on re-orientation

A continuation pack front-loads exactly the right context — task, decisions, files to look at — so the next session starts working immediately, not catching up.

With ContextOS

Lower cost per session

Less re-explanation means fewer input tokens per useful output. A tight continuation pack can replace a sprawling catch-up prompt at a fraction of the cost.

Try the shortest path first.

Install from PyPI, initialize project memory, and generate a continuation pack before wiring it into your real workflow.

# pip
pip install checkpoint-cli

# uv
uv add checkpoint-cli

checkpoint setup-user
checkpoint init
checkpoint status
checkpoint continue

Start the 5-minute quickstart Read how it works