Open Aura

Open Aura

Open-source, local-first: turn your real AI coding sessions into a scored builder profile, entirely on your own machine.

Open Aura is an MCP server your agent (Claude Code, Cursor, Claude Desktop, …) connects to. After a session, the agent sends a redacted evidence packet — truncated snippets + file metadata, never raw code or prompts — and Open Aura scores how you worked with the AI and stores it in your local Postgres. No login, no account, nothing leaves your machine.

Want a public profile, the shared leaderboard, and team insights? Those live in the hosted edition at vibelevel.ai.

---

Quickstart (Docker)

cp .env.example .env          # then set OPENAI_API_KEY to your OpenCode key
docker compose up             # Postgres + backend (MCP :8090) + viewer (:3000)
docker compose up -d --build  # rebuild images after pulling changes

Point your agent's MCP config at the local server (no auth in local mode):

// Claude Code / Cursor MCP config
{
  "mcpServers": {
    "aura": { "url": "http://localhost:8090/mcp" }
  }
}

Then, after a piece of work, ask your agent to "score this session with Aura."

Local viewer

A read-only web viewer ships with the stack (its own aura-ui container). After docker compose up, open http://localhost:3000. A left sidebar — Profile · Getting Started · Leaderboard · Sessions, a recent-sessions list, and a greyed Coming soon group (importing your sessions / submitting to the leaderboard — those live in the hosted edition) — sits beside the content: the hero score block, flip-able insight cards, dimension bars, and a read-only pull of the public leaderboard. It reads your local data through the backend's REST API and never scores or writes anything.

Updating / rebuilding

The stack bakes the source into its images at build time (aura-app from ., aura-ui via next build) — there are no live source mounts. So after you git pull new code, a plain docker compose up keeps running the old cached images and you won't see the changes. Rebuild:

docker compose up -d --build

If a change still isn't showing (a stale layer cache), force a clean rebuild:

docker compose build --no-cache aura-app aura-ui
docker compose up -d

Then hard-refresh the viewer (Ctrl/Cmd+Shift+R). The backend startup log prints the running version + build (e.g. Open Aura v0.2.0 · build …) so you can confirm the new code is live.

What it measures

A referenceless read of your process — no rubric, no test cases — across dimensions like prompting effectiveness, AI collaboration/steering, problem decomposition, and human contribution vs. AI reliance, plus a session archetype, insight cards, and measured telemetry (tokens, tools, "ships-it" lifecycle). Tools exposed: score_this_session, import_history, get_my_profile, whoami.

Privacy

The redaction contract is the whole point: raw transcripts and file contents never leave your machine. The agent sends only truncated text excerpts and file paths/metadata. In local mode the scoring and profile paths make no outbound calls — everything stays in your local Postgres. (The only exception is the viewer's optional Leaderboard tab, which does a plain read-only GET of the public leaderboard from vibelevel.ai — it sends none of your data.)

Configuration

Var	Default	Purpose
AURA_LOCAL_MODE	true (in .env.example)	single local user, no PAT/login
AURA_SCORING_MODEL	openai/gpt-oss-120b	scoring model (provider routing in model_config.json)
OPENAI_API_KEY / GROQ_API_KEY / ANTHROPIC_API_KEY	—	bring your own; set the one your model uses
OPENAI_BASE_URL	OpenAI default	optional OpenAI-compatible endpoint (for OpenCode, use https://opencode.ai/zen/v1)
POSTGRES_URL	compose-provided	local Postgres (apply schema.sql for bare-metal)
AURA_MCP_PORT	8090	MCP server port

Experimental: PFG operational insights

An opt-in, off-by-default capability that turns a scored session into concrete, tool-aware tips. The graph already maps the tools builders actually use — Claude skills, LangChain, the OpenAI SDK, MCP servers, and more — so Open Aura can look at what you built and nudge you toward them: "you hand-rolled this, but there's an established SDK for it," or "there's a Claude skill for exactly this." These check-tips are advisory only — they never change your Aura score.

It works precisely because Open Aura is local — it reads your real git diff and full transcript (passed as local_context, never persisted or scored) instead of a redacted summary.

To turn it on you need a read-only Personal Access Token from graph.vibelevel.ai — see the Getting access section in docs/PFG_INSIGHTS_POC.md and the PFG_* vars in .env.example.

How it's built

• src/aura_mcp/ — the MCP server (server.py) + the single-local-user binding (local_auth.py).
• src/services/aura/ — the referenceless scorer, signal extraction, archetypes, profile aggregation, and the scoring rubric (aura_model_coding.py / aura_model_writing.py) — open for contributions.
• src/core/ — slim config, Postgres pool, and LLM provider routing.
• aura_mcp_app.py — the entrypoint that mounts the MCP app + the REST API + /health.
• packages/aura-ui/ + web/ — the React viewer components and the Next.js host that renders them (the aura-ui container).

Roadmap

Current focus areas — contributions welcome (granular items live in Issues):

• Improve scoring — make the rubric more discriminating and harder to game: sharper archetype / human-contribution bands, a human-contribution score cap, and better-calibrated dimension scores.
• Import sessions to VibeLevel Aura — a browser-mediated export → sign-in → publish flow so your local Aura becomes a shareable, recruiter-facing profile on vibelevel.ai (nothing is uploaded automatically).
• Improve VibeGraph / PFG insights — tighter tag extraction and graph resolution (fewer loose matches), bounded/latency-safe grounding, and clear transparency about what stays local vs. what your BYO model sees.

Contributing

Issues and pull requests are welcome — see CONTRIBUTING.md.

License

Apache-2.0 — see LICENSE.

VibeLevel and Aura are names and brands of VibeLevel. This repository is the local-first, open edition. The hosted service at vibelevel.ai — the public leaderboard, shareable profiles, and the aggregated network — is separate and not included here. The license grants no rights to the VibeLevel or Aura names or logos, or to operate a hosted service using them.