open traces

open traces is a local-first evidence layer for agent work. It treats the session, rather than the diff, as the unit of work, and records what an agent sees, does, and changes. The product is framed around agent traces as a way to understand work produced by AI agents and the outcomes that follow.

Its CLI captures traces and supports setup, buckets, traces, trails, context trees, workflows, datasets, and security tools. The page describes a private trace bucket in which capture-time envelopes, patch history, trail and context companions, blobs, and manifests stay local, with content-hash deduplication to make republishing and machine switches safe. It also includes trace discovery through deterministic packets over a local lexical and concept index, Git-linked lineage features that connect trace patches to the history that accepted them, context-tree reconstruction that shows what the agent saw at a step and produces resume packets, and trace intelligence that can derive signals such as context waste, run health, and comparisons between two traces.

The workflow and review features turn raw trace evidence into schema-valid dataset rows for a chosen objective, with review actions to approve, reject, reset, schedule, or publish projected rows while raw traces remain in the bucket. Security tools include nine detectors, transformers, and a judge that run in one fixed order before egress and are explicit and off by default. The page also says the command-line interface emits structured JSON with next_steps on every command, and it is built for agents to drive agents. A web companion called the Hub makes shared buckets, trails, context trees, and datasets browsable, with trace intelligence on top. The page also refers to a hub preview with early access for individuals and teams.