Aura

Aura

Aura is a local-first AI agent framework for developers who want approval-gated execution, auditable state, and resumable workflows without giving up modern tool calling or model flexibility.

Built on top of Agno, Aura focuses on the part most agent demos skip: making agent behavior inspectable, replayable, and safe enough to use in real project directories.

Why Aura stands out

• Local-first project state: sessions, runs, approvals, and artifacts live under .aura/, so workflows stay versionable and reproducible.
• Approval gates for risky tools: read-only operations can flow fast, while shell or sensitive actions stop for explicit approval.
• Append-only audit trail: every model response, tool call, and approval event is persisted as JSONL for replay and debugging.
• Resumable runs: interrupted tool loops can pause to disk and continue later, including cross-process recovery.
• Multi-surface runtime: the same engine can back a CLI today and web, cloud, or editor surfaces later.
• Extensible orchestration layer: built-in tools, MCP servers, skills, DAG execution, and delegated subagents share one runtime model.

Architecture

flowchart LR
    Dev((Developer)) --> CLI["CLI surface<br/>aura/cli.py"]
    Dev --> Future["Future surfaces<br/>web / cloud / plugin"]

    CLI --> Engine["Async runtime engine<br/>aura/runtime/engine_agno_async.py"]
    Future --> Engine

    Engine --> Router["Model router + provider adapters"]
    Engine --> Tools["Tool runtime<br/>built-ins / MCP / DAG / subagents"]
    Engine --> Approval["Approval policy + inspection"]
    Engine --> Storage["Session, artifact, event stores"]

    Tools --> Workspace["Project workspace"]
    Approval --> Runs[".aura/runs/"]
    Storage --> Sessions[".aura/sessions/"]
    Storage --> Events[".aura/events/*.jsonl"]
    Storage --> Artifacts[".aura/artifacts/"]

Quick start

git clone https://github.com/qWaitCrypto/Aura.git
cd Aura

python -m venv .venv
source .venv/bin/activate
python -m pip install --upgrade pip
python -m pip install -e ".[dev]"

aura init demo-workspace

Edit demo-workspace/.aura/config/models.json and set the provider credentials and model you want to use, then start a session:

cd demo-workspace
aura chat

Run the integration suite:

python -m pytest -q -p no:cacheprovider scripts/integration

Core capabilities

Capability	What Aura adds
Tool execution	Approval-aware execution, inspection previews, and persisted tool results
Session memory	Event replay, compaction, and resumable runs stored on disk
Multi-model routing	OpenAI-compatible, OpenAI Codex, Anthropic, and Gemini adapters
Delegation	Subagents and DAG-based work orchestration with approval passthrough
Extensibility	Skills, MCP integration, optional knowledge retrieval, surface abstraction

Repository layout

aura/
├── cli.py                     # CLI entry point
├── runtime/                   # Engine, routing, tools, storage, orchestration
├── surfaces/                  # Surface stubs for future web/cloud adapters
└── ui/                        # Console rendering

scripts/
├── integration/               # Integration tests
├── mcp_smoke_server.py        # Local MCP smoke server
├── mcp_smoke_client.py        # MCP smoke client
└── smoke_agno_engine.py       # End-to-end smoke runner

Development

See CONTRIBUTING.md for setup, testing, and commit conventions.

Project status

Aura is currently CLI-first and actively evolving. The runtime already includes:

• async tool loops
• approval-aware execution
• MCP tool loading
• resumable runs
• delegated subagents
• DAG execution primitives

The main focus now is polishing the open-source surface area: packaging, CI, docs, and contributor ergonomics.

License

MIT