Questions & answers
What ORG-2 records, which agents it runs, where your code lives, and how it sits alongside the tools you already use. If something here is missing, ask in the issue tracker.
ORG-2 is an open-source desktop workspace that runs AI coding agents and records every session as a replayable trajectory. It supports 20+ coding agents including Claude Code, Codex, Cursor CLI, GitHub Copilot, Gemini, OpenCode, Aider, Goose, and Cline, and it keeps one shared record underneath all of them so a team can replay any session, review it, and trace a shipped line back to the decision that produced it.
Yes. ORG-2 is free to download and open source under the AGPL-3.0-or-later licence, and has been since June 15. The desktop app is a full product rather than a trial: running agents, recording sessions, and replaying them on your own machine are free and unmetered, and self-hosting the team backend on your own Supabase project is free too. Paid plans exist only for managed cloud sync, where ORG-2 runs that backend for you.
ORG-2 runs on macOS (Apple Silicon, .dmg), Windows (x64 .exe installer or .msi), and Linux (x64 .AppImage or .deb). It is built with Rust and Tauri rather than Electron, so the install is under 100MB on disk and uses a fraction of the RAM of a bundled-Chromium app you leave running all day.
No — ORG-2 runs those tools rather than replacing them. Claude Code, Cursor CLI, and Codex each record only their own sessions, so a team running three to five agents has no single place where that work is visible. ORG-2 sits underneath them as a neutral record: you launch the agents you already use, and every session lands in one timeline regardless of which tool produced it. ORG-2 also ships its own native Rust harness for people who want one, but using it is optional.
Because an agent vendor can never be the neutral record across its competitors: Claude Code will never record what Codex did. Per-tool history also stops at that tool's boundary, so it cannot answer cross-cutting questions like which agent touched a file, what a run cost across runtimes, or what a teammate's session actually did. ORG-2 ingests 15+ session formats into one record, which is what makes those questions answerable.
Git blame stops at who committed a line, and a pull request shows the diff after the reasoning is gone. ORG-2 records the trajectory that produced the change — the prompts, commands, file reads, edits, MCP tool calls, and browser use — so you can trace any shipped line back to the agent session and the decisions behind it. ORG-2 calls this AI blame. At agent speed the diff alone stops being reviewable: code written on Monday is legacy by Friday.
ORG-2 replays an agent session like a video, with a draggable scrubber and speeds from 1x to 6x. Six event types are recorded on one synchronized timeline: commands, file reads, file edits, MCP tool calls, skill loads, and browser use. You can also watch a session live as it runs, the way you would peer-code over a screen share.
Yes. On first run ORG-2 parses each supported CLI's on-disk session history and backfills it, so months of past sessions appear immediately instead of the record starting empty. After that, sessions from other tools are ingested in real time, including work that never touched the ORG-2 window.
ORG-2 runs and records 20+ coding agents, including Claude Code, Codex, Cursor CLI, GitHub Copilot, Gemini, Antigravity, Kiro CLI, OpenCode, Kimi Code, Aider, Goose, Amp, Cline, Kilo Code, Grok CLI, Devin, Qwen Code, Continue, Droid, and Mistral Vibe, alongside its own native harness. Sessions from 15+ of those formats can be replayed on the shared timeline.
ORG-2 is local-first: sessions are recorded and replayed on your own device, and using the app solo requires no account and no server. Team collaboration is opt-in and runs on Supabase — either a project you host yourself, which keeps shared session data entirely on infrastructure you control, or ORG-2's managed cloud if you would rather not operate one. Permissions are per-session at three levels: view, comment, and edit.
No. ORG-2 takes 0% margin on model usage — you connect your existing API keys and agent subscriptions and are billed directly by those providers. The app aggregates token spend and cost across every runtime on the device so you can see what your agents actually cost, but it never sits between you and the provider.
Yes, in part. Through a research partnership with OpenBMB, on-device models power proactive housekeeping agents inside the app, and the same architecture is the groundwork for fully on-premises deployments. Cloud and local runtimes roll up into one usage ledger either way.
Teams review the trajectory rather than only the diff. Comments anchor to a specific step of a session — an edit, a command, a decision — instead of to a line in a diff that has already lost the reasoning. A teammate's run is one click from replay, and any session can be forked to continue with full context in your own workspace.
Agent runs are cards on a board with five columns: Todo, In Progress, Blocking, Turn Finished, and Archived. Each card carries the model that ran, its diff stats, who started it, and how long ago. Runs that need human sign-off surface in the Blocking column rather than dying silently in a closed terminal, and finished runs auto-archive after 24 hours so the board stays current without manual upkeep.
Drift detection flags the moment an agent's execution diverges from what was asked. Because the record holds the goal, what the agent understood, and what it actually did, hooks can compare intent against execution and surface malpractice while a run is still in flight rather than at review time.