Atomic

Atomic is a source code management system built as a distributed, agent-native alternative to Git. Rather than tracking file-level snapshots, it operates at the primitive level and links semantic call graphs and code search to intents, acceptance criteria, tasks, changes, evidence, provenance, attestations, reviews, views, and memories inside a single causal graph that is shared across the AI coding agents a team already uses. Stable change identities preserve those relationships when work is combined or reorganized, so the graph itself functions as the state agents operate from rather than an audit trail reconstructed afterwards.

The product is delivered as a single command-line interface. Repositories are enriched with AST analysis, tree-sitter entity extraction, and a semantic graph addressable at the function, class, and token level, allowing agents to query relationships such as callers of a specific entity or neighbors of an intent instead of reading files. Commands include triage review workflows for merging branches and vault queries against the graph. Atomic also captures the full causal graph of each agent session — what was requested, explored, read, and modified, along with the model used, turn count, and duration — and binds that record to an immutable change identity that travels with the change through squash, rebase, or cherry-pick operations. Sessions are ed25519-signed and content-addressed, and the resulting history can be queried by intent, model, file, or decision.

Atomic is installed via a shell script from atomic.storage, runs on macOS, Linux, and WSL2, is approximately 14 MB, and is offered under the Apache 2.0 license. Installation requires no account and the tool ships with no telemetry enabled by default; the installer detects the active agent and begins indexing the current repository.

Business model

Atomic distributes an Apache 2.0-licensed command-line tool that installs directly on a developer machine without requiring an account. Pricing or paid offerings are not described.

Traction

Published performance figures cite 44% faster agent sessions and 82% less context read relative to baseline agents on the same task set, with one graph query replacing five standard tool turns.