Aura Code Audit Trail
Track what coding agents changed and why, then block commits whose symbol-level changes violate their stated intent.
- Source repo
- Naridon-Inc/aura
- Stars
- ★ 49
- Last updated
- 27d ago
- License
- Apache-2.0
- Primary language
- Rust
- FA score
- 50/100 · Major gaps
At a glance
- How it runs
- Works with
- Universal · cross-platformCodex · Claude Code
- Cost
- Free, no paid service needed
- Setup effort
- Low · running in minutes
- You'll need
- Typical use
- A team running several coding agents needs to determine months later which agent changed a function, what request initiated the work, and what intent was recorded.
- Not a fit if
- Teams auditing codebases outside the 13 supported languages
- Teams unwilling to add .aura/ data, Git notes and commit hooks to their repo
- Enterprises needing documented cloud sync deployment and operations
- Source review
- 50/100 · Major gaps
What does this agent do, and when should you use it?
Aura adds a semantic audit and versioning layer on top of Git, recording which symbols an agent or developer changed, the stated reason, and signed provenance for each commit. A single Rust engine powers the aura CLI, a Tauri and React desktop application, and a VS Code extension, all of which share the same .aura/ meaning plane. Git hooks parse each commit with Tree-sitter, represent functions and classes in an AST Merkle graph with call edges, and compare the resulting changes with the declared intent. Its outputs include queryable provenance, semantic-review findings, signed metadata, AI usage and cost records, and function-level rewinds that do not revert the surrounding commit. Data stays in .aura/ and Git notes by default, requires no server, and leaves the machine only when cloud sync is configured.
After aura init installs Git hooks, Aura parses every commit into logic nodes, computes AST-level differences, and links those differences to the intent recorded while the commit is being written. Its pre-commit gatekeeper can reject a commit when the stated task and the affected code disagree, such as an intent concerning login that modifies billing. aura why <file> reports the landed commit, author or agent, originating request, recorded intent, and capture method for a file or line. The CLI also exposes semantic review, AST-node-based function rewind, the Crew work loop, AI usage and cost tracking, and an MCP server started with aura mcp; the desktop and VS Code interfaces use the same .aura/ records. Tree-sitter analysis is documented for Rust, Python, TypeScript, JavaScript, Go, Java, C#, C++, C, Ruby, PHP, Swift, and Kotlin.
- A team running several coding agents needs to determine months later which agent changed a function, what request initiated the work, and what intent was recorded.
- An engineering organization with audit obligations needs signed, symbol-level provenance and a stated reason for code changes.
- A maintainer wants to stop out-of-scope edits before commit, such as an agent deleting authentication logic while claiming to tune retry behavior.
- A developer needs to restore one faulty function without reverting unrelated work contained in the same commit.
- A large-codebase review team wants to detect layer violations, silent deletions, and architectural drift from graph and AST changes rather than text alone.
- A team using Claude Code, Gemini CLI, Codex CLI, Kimi, Cursor, or OpenCode wants repository-local records covering changes produced by several tools.
How do you install or deploy this agent?
The prebuilt installer is documented for macOS, Linux, and Windows and requires a shell capable of running the Bash command:
curl -fsSL https://auravcs.com/install.sh | bash
cd your-repo && aura initaura init installs hooks in the current Git repository. To build from source, install Rust and Cargo, then run this from the repository root:
cargo build --releaseNo server or cloud credentials are required for the local workflow. The supplied material does not document the account, credential, or configuration requirements for optional cloud sync.
How do you use this agent?
Once initialized, create commits normally; the hooks perform AST parsing, semantic diffing, intent logging, and gate checks. Query the reason behind changes to a file with:
aura why src/billing.rsFor direct access from an MCP-capable coding agent, add this server definition to the client's MCP configuration:
{
"mcpServers": {
"aura-vcs": { "command": "aura", "args": ["mcp"] }
}
}The MCP server exposes more than 30 tools, including aura_usage for agent spend monitoring. Telemetry can be disabled with either documented environment variable:
export AURA_TELEMETRY_OPTOUT=1export DO_NOT_TRACK=1What are this agent's strengths and limitations?
- It records intent, provenance, and signatures at function and class level instead of limiting history to changed text lines.
- The pre-commit gatekeeper compares declared intent with actual AST changes and can stop task drift before a commit lands.
- Function-level rewind operates on AST nodes, allowing one function to be restored without reverting neighboring work or creating text merge conflicts for that operation.
- The CLI, desktop application, and VS Code extension share one
.aura/meaning plane, so records remain consistent across interfaces. - Core records live in
.aura/and Git notes, allowing local operation without deploying a server.
- Adoption adds
.aura/data, Git notes, and commit hooks to the repository workflow, all of which the team must account for operationally. - Semantic parsing is documented for 13 languages; the source provides no evidence that other languages receive equivalent AST-level auditing or rewind support.
- Integration depth varies by client: Claude Code has MCP tools, a status line, and transcript parsing, while Cursor is only explicitly associated with workspace detection.
- Cloud sync is mentioned but its deployment, credentials, pricing, and failure handling are not documented in the supplied material.
- Source builds require Rust and Cargo, while the desktop application and VS Code extension have their own component-specific build locations.
How does this agent compare with similar options?
Aura complements rather than replaces Git. Git continues to manage bytes, text diffs, commits, branches, history, and remotes; Aura adds an AST Merkle graph, logic nodes, call edges, intent, provenance, and signed metadata. Compared with the line ownership shown by git blame, Aura is designed to identify the developer or agent responsible for a symbol-level change, explain why it was made, and check whether the implementation matched its declared intent.
Key facts side by side with the most closely related agents.
| Agent | Source review | Form / cost | Stars | Updated | Language | Full support on |
|---|---|---|---|---|---|---|
| Aura Code Audit Trail This agent | 50 · Major gaps | CLIFree | ★ 49 | 27d ago | Rust | Codex · Claude Code |
| Argot Repository Style Analyzer | 78 · Good | CLIFree | ★ 49 | 2mo ago | Rust | Claude Code |
| Gortex Code Intelligence | 73 · Some gaps | CLIFree | ★ 2k | today | Go | Codex · Claude Code · OpenAI API · Claude API |
| git-lrc Commit-Time Code Review | 66 · Some gaps | CLIFreemium | ★ 1.5k | 16d ago | Go | Claude Code · OpenAI API · Claude API |
How does FollowAgents rate this agent?
Why each dimension lost points
The README says data stays local unless cloud sync is configured and documents telemetry opt-outs; the source shows telemetry deriving a stable hashed machine identifier from hostname and username and sending event metadata. The sample does not establish the full MCP/cloud permission boundary or fully specify telemetry fields and retention, so least privilege, sensitive-data handling, and external effects receive partial credit. Installation configures Git hooks, and the docs describe intent gating and function-level recovery, but the sample does not show that all writes require confirmation or that recovery is always conflict-free. Source attribution and signed metadata are described, but the verification flow is not fully shown.
The README, CLAUDE.md, and module list describe a broad CLI, desktop, extension, and MCP product, but CLAUDE.md's mandatory workflow does not fully align with the README's short usage path; team sync and automatic-push claims are not corroborated by the provided code, reducing self-consistency. The Cargo workspace and local crate paths support the fresh-clone build claim and list many versioned dependencies, though installer and runtime dependency availability remain underspecified. The daemon entry point reports path, configuration, and service errors with messages and exit codes, which provides partial evidence for failure handling.
The docs cover human commits and multiple coding agents, and enumerate languages, CLI, desktop, and VS Code scenarios. Privacy guidance identifies cloud sync as an exception for outbound data, but the sample does not explain permission and capability boundaries for the 30+ MCP tools, cloud features, hooks, and desktop actions. README gives installation, initialization, and MCP configuration; CLAUDE.md and the Gemini skill specify triggers, though the skill packet is only a sample and one skill requires writing an intent file, leaving overall applicability unclear. Linux/macOS CI provides partial support for cross-platform claims, while the supplied release matrix omits Windows.
README is organized around quick start, problem, operation, product surfaces, integrations, privacy, and repository layout, with guide, contribution, and security-policy links. Installation guidance is brief, and the curl-to-shell command is not accompanied in the sample by verification, uninstall, or platform-specific steps. The product name is stable, but README says 30+ tools, CLAUDE.md says 29+, and its stated version 0.12.6 conflicts with Cargo's 0.19.58. FAQ and integration examples are useful; SECURITY.md documents pre-1.0 support for only the latest release and service scope. The Apache 2.0 license text is complete. Release badges, version metadata, and workflow provide some versioning evidence, but no changelog content is supplied. SECURITY.md describes a small maintenance team and private reporting path; publisher identity remains unknown and earns no brand credit.
The product aims to record what agents changed and why, with intent gating, semantic review, and function-level recovery; README includes a query output example and MCP configuration, giving users a concrete output shape. If the described capabilities work, they add agent attribution and intent beyond Git blame, but the static sample does not establish core detection accuracy. The CLI, desktop app, extension, cloud sync, and multi-agent features create substantial setup and learning costs; quick-start and integration documentation partly offset them.
The README's local-storage, telemetry opt-out, and cross-platform CI claims have partial counterparts in the source and workflow samples. Core claims about intent enforcement, signatures, and conflict-free recovery are not sufficiently corroborated by the supplied source files. The packet distinguishes product descriptions, explicit security policy, and visible implementation excerpts, but several README assertions remain unverified against code, so only partial credit is warranted.
- The install command pipes a remote script directly to a shell; verification and uninstall steps are not shown in the sample.
- Telemetry uses a stable hash derived from hostname and username; review actual event contents and cloud-sync configuration before deployment.
- Tool counts and version references differ across README, CLAUDE.md, and Cargo.toml; some key safety and recovery claims are not corroborated by the supplied code.
Maintain this project? Add the FARS badge to your README
[](https://followagents.com/en/agents/aura)<a href="https://followagents.com/en/agents/aura"><img src="https://followagents.com/badge/aura.svg" alt="FARS review"></a>Recent releases
Latest entries from the project's GitHub Releases. Notes are quoted from upstream as written, not rewritten or tested by FollowAgents.
-
v0.19.58 v0.19.58 — in front, and fewer places to look
Release notes excerpt
This release catches the public tree up with four desktop releases, 0.19.55 through 0.19.58. Panels stopped opening behind the terminal. The terminal is drawn by the graphics card, which makes it a layer the rest of the window cannot stack against — so nine different popovers had been opening underneath it, and clicking one gave you a sliver of a panel instead of a panel. The one people hit most was the small "old aura on PATH" chip in the status bar. …View on GitHub ↗ -
v0.19.54 v0.19.54 — Share a session, they open your files
Release notes excerpt
Share a session and the other person can now open your project's files from their own machine — no clone, no branch, no git setup at all. They see what you see, including work you haven't committed yet. Long conversations also stay quick: a chat used to get slower the longer it ran, and no longer does. Three fixes as well — the app no longer goes blank when you hover a file an agent linked, Chat now shows when a terminal agent is still working instead of looking finished, and a full disk no longer hides the reason the app quit. …View on GitHub ↗ -
v0.19.47 v0.19.47 — Aura stops overstating what it knows
Release notes excerpt
Aura stops overstating what it knows The honest-claims release. Several surfaces used to present a guess with the same confidence as a fact — a file whose reason was inferred from the commit message read exactly like one whose reason was logged by the agent that made the change. They now say which one you are looking at, and say nothing where they have nothing. That sounds small. It is the whole product: an audit trail that pads its gaps is not an audit trail. In this release …View on GitHub ↗
FAQ
Does Aura send code or audit records to an external service?
.aura/ storage and Git notes, and data leaves the machine only if cloud sync is configured. Usage tracking reads local Claude Code transcripts and makes no API calls.Will it replace or rewrite the existing Git workflow?
.aura/, and Git notes.Is Claude Code required?
Do intent checks and rewind work for human-written commits?
Does local use require a cloud account or persistent server?
Thanks — we use this to improve the page.