Claude Observatory

One change at a time

A change accepted without being understood is a liability. A large diff reviewed only at the end of a session is understood too late to steer — by then the code is already in.

Existing tools address parts of this: Usage meters count tokens; log viewers replay a session; editor assistants and Claude Code's own /rewind roll back to a checkpoint; orchestrators launch fleets of agents and review each branch as a diff to merge. A checkpoint is coarse, though: it reverts a whole turn rather than one edit, does not track the files a shell command changed, and clears when the session does. None keeps a finer, longer-lived record you can act on one change at a time — accept or reverse any single edit while the rest stand, including the files a Bash command touched, from the terminal or either editor. That is the record Claude Observatory keeps.

Outside the model loop

Capture runs in local hooks, outside the model loop: it costs no tokens and changes nothing about what the agent sees or the decisions it makes. The record builds itself as the agent works, so it is there before you need it — not something you reconstruct after the context is gone. Everything stays on your machine; nothing is sent anywhere.

From one edit to the whole trace

• Any edit, reversible on its own. Undo a single change while the edits around it stay — a position-anchored merge preserves them, and when a later edit genuinely overlaps, the undo is refused rather than applied blindly. No clean git state required, and never a silent clobber.
• Review by the unit of work, not just the edit. Follow a whole subtask — its edits in the order the agent made them, across files — as one coherent sequence, and scale the same review up to a chapter, a task, or a workflow run. Single edits when you want precision; traces when you want the shape of the work.
• Every edit, recorded. Each text edit is saved the moment an agent makes it — including the files a shell command changed. Anything it cannot snapshot, such as a binary or an oversized file, is marked as an explicit skip, never dropped silently.
• Cross-platform support. VS Code, JetBrains IDEs, and the terminal read the same local store, so review never depends on a particular editor or an outside service.
• Every tool call, explained. Reasoning, a timeline of each tool call, and flags of risky shell commands and off-machine network access — read from the transcript the session already produced.
• Many agents, one view. When several agents run in parallel across your git worktrees, it unifies them — who is working, on what, and where their changes collide.

License & credits

Claude Observatory is developed in the open by the Cell Observatory project, licensed under Apache-2.0, with a Contributor Covenant code of conduct. It is built with and for Claude Code. Adjacent work informed its direction: claude-devtools on reading the on-disk transcript, ccusage and sniffly on local-first analytics, and the worktree-fleet tools (Conductor, claude-squad, vibe-kanban) on running agents in parallel — where Observatory adds the layer that observes them.