Supafork gives your agent sessions a GitHub-style home
I've been running Claude Code, Codex, Gemini CLI, OpenCode, Pi, and a rotating pile of coding harnesses for months. Every one of them stores sessions locally. None of them talk to each other. So if I want to reuse a clever prompt chain I built in Claude Code inside Codex, I'm copy-pasting markdown. If I want to hand a working session to a coworker, I export a JSON blob and hope their harness version matches. That's the gap Supafork fills.
The pitch is basically: GitHub for agent sessions. Save a session, drop it into a collection, flip the visibility to public or private, done. The two features that actually matter:
- Forking sessions — grab someone else's session and run it against your own data. The session is the artifact now. This is the part that turns a personal log into a reusable workflow.
- Cross-harness import — pull a session from Claude Code and load it into Codex, or whatever else supports the import format. Not every pairing works perfectly (the author is upfront about harness coverage being uneven), but the workflow is real.
The CLI is the entry point. Install, run your agent, and it pipes the session up to beta.supafork.com automatically. There's a web UI for browsing, forking, and organizing into collections. Sharing is a URL, not a screenshot.
A few things worth flagging if you're thinking about using it:
1. Version drift is the silent killer. Agent harnesses change their session schema between releases. A session that imports cleanly today might break next month. Ask how the project handles migrations before you bet a team workflow on it.
2. Privacy defaults matter. Sessions can contain proprietary code, API keys, internal URLs. The platform needs sane defaults (private by default, explicit sharing) and good redaction tooling. Worth testing with a throwaway repo before you upload anything sensitive.
3. Search beats browsing once your library grows. Collections are nice at 20 sessions. At 200, you want full-text search across prompts, tool calls, and outputs. GitHub got away with weak search for years; agents won't.
The fork semantics are what I'm watching most closely. If a fork can diverge cleanly — new branch metaphor, lineage tracking, the ability to pull improvements back upstream — then this becomes genuine infrastructure for the agent ecosystem. If fork just means "copy and rename," it's a glorified pastebin.
Try it at beta.supafork.com. The CLI install is one command and the import path for Claude Code and Codex already feels solid in my testing. Less smooth on some of the smaller harnesses, but that's expected for an early build.
What's broken for you when you try to share an agent session today? Is it the format mismatch, the lack of versioning, or just that nobody on your team knows which harness you used?