Open Conversation Format
Your AI conversations are scattered. Claude has them in ~/.claude/projects/. Cursor in a SQLite blob you can't read with sqlite3 alone. Codex in ~/.codex/sessions/. ChatGPT in a tree-shaped JSON tarball that linearizes lossy. None of them talk to each other. None of them stay attached to your code when you commit.
So a few things happen - repeatedly:
- Your Claude budget hits zero mid-project. You finish in Cursor. The conversation context doesn't follow.
- Your colleague reviews a PR. They see the code you and the AI agreed on, not the four wrong tries before it.
- You check out a six-month-old commit. The AI session that produced it is gone, archived, or trapped in a dead subscription.
- You switch machines. Your ~/.claude/ is back at the office. The conversation that's been guiding this feature is unreachable.
Open Conversation Format (OCF) is one file format that holds an AI conversation: messages, tool calls, file edits, terminal commands, attachments, code blocks, model and token telemetry. Wire-compatible with OpenAI Chat Completions for the message core. Session envelope around it for everything OpenAI's wire doesn't cover.
The point isn't a new chat protocol or a new agent runtime. The point is portability - conversations you can move between tools, attach to commits, review in PRs, resume from any checkpoint, hand to a colleague.
| Repo | What's there | Status |
|---|---|---|
| spec | The format. JSON Schema, platform mappings, tool conventions, reference examples. Start here. | Live - v0.1.0-draft |
| ocf-py | Python implementation: schema validator, canonical JSON, exporters for Codex / Claude Code / Cursor, Markdown renderer, ocf CLI. | Live - 0.1.0.dev0 (pre-alpha) |
| ocf-precommit | Git hook that snapshots active AI sessions into .ocf/ on every commit, so your repo carries both code and reasoning. | Planned |
| importers | OCF → Claude Code / Cursor / Codex. For switching tools mid-project without losing context. | Planned |
Empty repos are not pre-created. They appear when there's actual content.
Format standards live or die on whether they solve a real pain. The pain here: AI conversations are ephemeral, code isn't. Versioning code without versioning the reasoning that produced it has been the default for the entire AI-assisted-coding era. OCF is the smallest format that lets you stop doing that.
Read the ROADMAP in the spec repo for the build order.
- Spec: 0.1.0-draft. Schema may change incompatibly until 0.1.0 stable.
- Implementations: ocf-py is public (pre-alpha), validated against ~1400 real conversations across Codex / Claude Code / Cursor on the maintainer's machine. Other implementations land when there's something to ship.
- Contributing: spec issues and discussions go to the spec repo. Implementation issues will live in their respective repos once those exist.
Popular repositories Loading
Repositories
People
This organization has no public members. You must be a member to see who’s a part of this organization.
Top languages
Loading…
Most used topics
Loading…