An open-source tool called cross-agent-sync hit version 0.2.2 on npm this week — a few hundred lines of code that solve a problem every major AI vendor is dodging. Claude Code, Codex, Cursor, and OpenCode (all AI coding assistants) keep conversation records in incompatible silos; switching tools means restating your objectives, what's done, what's pending, and the pitfalls you've hit. We think this matters to working professionals — it tears a hole in the "AI does it all" narrative: today's AI tools can't even move "what was said" between each other.
What this is
Developer lhbDesign open-sourced cross-agent-sync on GitHub (CLI command: ass); it shipped v0.2.2 this week. The motivation is concrete: he was running four AI coding assistants on the same project and discovered that each tool stores "conversation history" in a completely different format — Claude and Codex use JSONL (a text-based log format), Cursor and OpenCode use SQLite (a lightweight database), and image attachments follow yet another scheme per vendor. Switching tools means restating objectives, what's done, what's pending, and the pitfalls you've hit.
The tool does four things: aggregate each agent's sessions into a read-only view, align them by code project root directory, stitch multi-segment sessions into "tasks," and generate handoff summaries (including images) that can be fed back into a new session. Once MCP (a protocol that lets AI call external tools) is wired in, a new session can "pull in an old conversation" to continue work. The design is strictly read-only, to avoid corrupting the user's existing session stores.
Worth noting: the tool itself was written through multi-agent collaboration — Claude handled the main code, Codex picked up where it left off, and Cursor polished the docs.
Industry view
On the upside, this validates a thesis we keep returning to: the real bottleneck for agents isn't model capability, it's engineering. This open-source tool fills a "cross-vendor glue layer" that giants like Anthropic, OpenAI, and Cursor have zero incentive to build. It cleanly separates the "session vs. task" abstraction — a more disciplined model than what many big-company product managers ship.
But the counterarguments deserve equal airtime. First, an individual developer's side project is not an enterprise solution — SQLite read-only protections, cross-platform compatibility, and version compatibility are each deep rabbit holes; the author himself admits they "only surfaced after launch." Second, the industry doesn't actually want interoperability. Anthropic wants you using only Claude Code; OpenAI wants you using only Codex; the data walls are the moat. An open-source tool challenges the foundation of these business models — clicking "star" on GitHub is easy; getting big vendors to actually cooperate is nearly impossible. Third, the author deliberately made the tool "say so when it can't read something, not silently emit empty columns" — that honesty is rare in most AI tools, and it should remind us that a lot of "intelligent experience" is built on a compromise of "fake it when you can't guess."
Impact on regular people
For enterprise IT: if you're evaluating AI assistants for employees, recognize that the "switching cost" is far higher than it looks — this isn't like installing a new app; it's a total reset of historical context, collaboration records, and decision rationale.
For individual professionals: if you're still agonizing over "which AI is better," ask yourself one basic question first — for the work you handed to AI three months ago, can you still find that conversation and its reasoning? If not, what you're using isn't AI; it's an intern with amnesia.
For the consumer market: the subtext here is that the "universal AI assistant" won't arrive anytime soon. What's more likely to solve your actual problems is a combination — one good at retrieval, one at writing, one at code — rather than one super-app that's "a jack of all trades, master of none." This is a sobering rebuttal to every "one-stop AI" marketing pitch out there.