Aura: product and architecture
Aura is a version-control layer that sits on top of Git and tracks what AI coding agents actually changed, function by function rather than line by line. It is built in Rust, published under Apache-2.0 at github.com/Naridon-Inc/aura, and is in free public beta. This Aura review covers what the product does, how its pricing works, and where it fits, based on the vendor's own documentation and repository.
Overview
The problem Aura sets out to solve is stated plainly on its own site: "Every check passed. The change was still wrong." Tests, code review and Git all answer questions about the code in front of them. None of them knows which behaviour an agent was allowed to change, so none can tell a legitimate cleanup from a deletion nobody asked for.
Aura's answer is to record intent before the agent runs. You approve what an agent may touch; when a commit is attempted, Aura compares that approval against what is actually staged, at the level of functions and exports rather than diff lines. If an export has disappeared, it names it, traces who called it, restores that one function, and leaves the rest of the cleanup alone. The check runs through a commit hook as aura verify-intent --staged.
The positioning is aimed as much at non-engineers as engineers: "Trust the code you didn't write, even if you can't read a line of it." That framing explains the emphasis on plain-English explanations of diffs and on proof of authorship rather than on raw diff tooling.
Key Features and Architecture
Aura is a desktop application with a CLI and 29 MCP tools, backed by a semantic history of every change an agent makes. Four capabilities carry most of the product.
Semantic history and rewind. Changes are recorded per function and per symbol, so a single function can be rewound without reverting an entire commit. Every diff can be explained in plain language.
Provenance and proof. Each agent run produces a sealed record — the vendor calls it a Genuine record — with a block identifier, showing what was asked, what was changed and whether the result matched the request. On Enterprise this becomes a signed, exportable AI-authorship ledger, recorded per symbol.
Mission Control. One window drives the agents a team already uses: Claude Code, Gemini CLI, Antigravity CLI, Cursor Agents, Codex, Open Code, Kimi Code, Pi, and any MCP-speaking agent. They share one history, one composer and one project view, and you can switch model mid-thread.
Team Radar. For teams, Aura tracks who is working in which symbol and warns about collisions before they happen — "You're both editing sort.rs" — alongside a shared task board and Crew, which runs several agents against one dependency graph.
The engine is self-hostable: you can run your own git remote and mirror repositories onto a machine you own, and the vendor states the open-source build never phones home. Cloud sync is the paid boundary, metered per month by tier.
Ideal Use Cases
Teams whose code is largely agent-written. The product's stated centre of gravity is more than one person in the same code, with several agents running. Team Radar and the shared provenance ledger only matter at that point.
Non-engineers shipping real software. The plain-English diff explanations and per-function rewind are aimed at someone who can direct an agent but cannot audit its output line by line.
Organisations that will face an authorship question. Aura's Enterprise framing is explicit about this: "teams whose auditors will eventually ask who wrote this line." A signed, per-symbol authorship ledger is a compliance artefact, not a developer convenience.
Solo developers wanting an agent safety net. The free tier is the complete engine, so a single developer can use semantic history, rewind and intent verification at no cost, self-hosted.
It is a weaker fit where agents are not part of the workflow: the entire value rests on knowing what an agent changed and whether that matched the request. A team writing code by hand gets little from it that Git does not already provide.
Strengths & Trade-offs
Strengths
- The full engine is open source under Apache-2.0, not a limited free tier, and can be self-hosted with your own git remote.
- Intent verification catches a class of defect that tests, review and Git structurally cannot: a change that passes every check but was never requested.
- Per-function rewind is finer-grained than a commit revert.
- One interface drives nine named agent CLIs plus any MCP agent, on shared history.
- Seat pricing is decoupled from agent count, and bringing your own model key is free on every tier.
Limitations
- The project is young and small: 44 GitHub stars, 2 forks, and a free public beta, so it carries the adoption risk of any early product.
- Two advertised Team and Enterprise capabilities — Aura provisioning the box, and org-wide mirror policy with a fleet view — are marked "soon" on the vendor's own comparison table rather than shipping today.
- Cloud sync allowances are capped per tier and no overage rate is published.
- Model usage on Pro and Team is metered on top of the seat price, so the headline figure is not the whole bill.
- There is no annual discount; everything is billed monthly.