Carryover runs as a local daemon, capturing AI agent context from on-disk transcripts so your agents pick up where you left off — across Claude Code, Cursor, and Codex.
Three guarantees. Each one solves a failure mode that costs you a session.
Hooks fire locally. The daemon reads transcripts directly from disk. Your agent does no writing, no summarizing, no token-tax — the bookkeeping is out-of-band.
A structured handoff — task, next_action, progress_log — is the contract. Old state stays in a local SQLite ledger, never the prompt.
Claude Code, Cursor, and Codex converge on the same AGENTS.md rail. Switch tools mid-task. Hit a compaction event. Same handoff, same place.
No agent plugin. No prompt overhead. Carryover runs as carryoverd on localhost:47823.
One command installs the binary and provisions the local ledger at ~/.carryover. Auto-starts via systemd-user on Linux; launchd autostart on macOS lands in v0.2 — for now, run carryoverd start once after install.
carryoverd watches transcripts in the background. No flags, no wrappers, no changes to how you already work with your agent.
Open Cursor. Open Codex. Hit compaction. The next session reads the structured handoff from AGENTS.md automatically.
Carryover ships as a single static binary. No runtime dependency.
Then run carryoverd install for one-time setup.
Capture every session. Hand off cleanly. Switch tools without losing the plot.
