SuperAPI AI Directory

aura

Open-source semantic layer on top of Git for reviewing AI-generated code, tracking intent/provenance, and surfacing risky change areas.

Open website

About this tool

Aura is an open-source project developed by Naridon, Inc. in Switzerland, positioned as a semantic layer on top of Git to provide an audit trail for the changes AI agents make in a codebase and the reasons behind them. Its core description is: signed, recorded per commit, tracked at the symbol level. It does not replace Git, but overlays on top of it; the user's history, remotes, and workflows remain unchanged, and Aura writes to .aura/ as well as git notes.

Purpose and the problems it solves The material points out that when multiple agents commit code in the same repository, git blame only shows that the commit was made by you, and the diff only shows which lines moved, but cannot show which agent changed which lines, what it was trying to do, or whether it actually did what it claimed. Aura attempts to record this information. Four capabilities derive from it:

  1. Auditable provenance: records who or what changed each symbol, the stated reason, and signatures for both.
  2. Intent gating: a pre-commit hook compares the declared intent with the AST changes that actually occurred. For example, if a commit claims to adjust login but touches billing code, the commit is blocked before it lands.
  3. Function-level rollback: a single function can be restored to its last good state without reverting the entire commit around it; because it operates on AST nodes, it does not produce merge conflicts.
  4. Semantic review: by comparing graphs rather than text, it detects layer violations, silent deletions, and architectural drift.

Core idea The material describes Git as "versioning bytes," while Aura adds a second plane, namely "versioning meaning": a graph containing every function and class, the edges between them, and the intent and provenance behind each change. Both planes are stored in your repository, with no server required; unless cloud sync is configured, the data does not leave the local machine.

One engine, three interfaces The repository is the complete Aura, with a single Rust engine exposed in three ways. All three read and write the same .aura/ meaning plane, so a change made by an agent on the command line appears in the desktop app with full provenance information.

  • aura CLI: includes git hooks, semantic review, function-level rollback, AI usage and cost tracking, the Crew work loop, and an MCP server (src/ + crates/).
  • Desktop app: a native Agentic Development Environment that can run multiple coding agents in the same workspace, including the Crew panel,Semantic review, rollback, and project timelines, with support for macOS, Linux, and Windows (desktop/).
  • VS Code extension: provides agent sessions, review, and git interface within VS Code (extensions/vscode). Build from source: run cargo build --release at the repository root.

MCP and multi-agent support Aura exposes an MCP server for direct integration by AI agents, providing more than 30 tools, including aura_usage (agents can self-monitor their own spending). It detects and tracks sessions from the following sources:

  • Claude Code: MCP tools + status bar + transcript parsing
  • Gemini CLI: MCP server + hooks
  • Codex CLI: session and transcript parsing
  • Kimi (Moonshot): OpenAI-compatible endpoint
  • Cursor: workspace detection
  • OpenCode: environment variables and configuration The materials also mention integration guides for OpenCode (quick start via the git-hook collection path) and Cursor (semantic review and provenance tracking alongside Cursor). The git-hook collection path works with any tool that makes commits.

Supported languages 13 languages are supported via Tree-sitter: Rust, Python, TypeScript, JavaScript, Go, Java, C#, C++, C, Ruby, PHP, Swift, Kotlin.

Use cases

  • Need to provide auditors with provenance records for code changes produced by AI agents.
  • Multiple agents in a team committing at the same time, and need to locate the attribution and rationale for changes by symbol.
  • Want to prevent commits whose claims do not match the actual changes from landing through intent gating.
  • Need function-level rollback without causing merge conflicts.
  • Want semantic-level review beyond textual diffs.

Limitations and notes

  • Does not replace Git, only acts as a layer on top of it.
  • Does not send code off the local machine unless cloud sync is configured.
  • Telemetry can be disabled via AURA_TELEMETRY_OPTOUT=1 or DO_NOT_TRACK=1.
  • Usage tracking only reads local Claude Code transcripts and makes no API calls.
  • Even if no agents are used,intent gating, rollback, and semantic review can also be used for human commits, and the provenance will simply show as yourself.
  • The license is Apache License 2.0, Copyright (c) 2026 Naridon, Inc.
  • Every commit in the repository passes through these gates as described, issues marked up-for-grabs are scoped to one file or one command; CONTRIBUTING.md covers building, the three commands run by CI, and where in the tree to start.

Regarding pricing, charges, specific results, or other matters not appearing in the materials, none are provided at this time.