AEGIS

An AI advisory board for your codebase that argues — and fact-checks itself against your repo. Seven expert advisors assess architecture, product, design, finance, security, privacy, and strategy, then their factual claims are verified deterministically with git and grep, not by another model.

What It Does

AEGIS adds a virtual advisory board to your development workflow. Each advisor has distinct biases, frameworks, and pushback triggers. They disagree with each other productively, a Dissent Index quantifies how split they are, and a recommendation engine synthesizes their perspectives into actionable recommendations. Then comes the part a plain LLM chat can't do: AEGIS extracts the advisors' factual claims and checks them against your actual repo — cited files must exist, cited code must say what the advisor said — flagging refuted claims with the exact command used as evidence.

• The Architect — Technical architecture, deploy speed, blast radius
• The Strategist — Product, user impact, market fit, prioritization
• The Designer — Design quality, UX, accessibility
• The Analyst — Cost, revenue impact, resource efficiency
• The Sentinel — Security, compliance, threat modeling
• The Guardian — Privacy, data handling, PII risk
• The Advisor — Startup-advisor cross-cutting voice (CLI-only)

Six advisors render in the dashboard grid; the seventh (The Advisor) is available via the CLI and MCP server.

Try It In 30 Seconds

No install. Run it directly from npm:

npx aegis-ai dashboard

Your browser opens to http://localhost:18789. Press Ctrl+C to stop.

Install Persistently

npm install -g aegis-ai

Then aegis dashboard works from anywhere.

Connect to Claude Code (optional)

Give Claude Code governance tools:

npm install -g aegis-ai-mcp
claude mcp add --transport stdio --scope user aegis -- aegis-mcp

Quick Start

# Initialize governance in your project
cd your-project
aegis init

# Launch the dashboard
aegis dashboard

# Or use the CLI directly
aegis status
aegis retro

CI Gate: fail the build on refuted AI claims

aegis verify re-checks recorded AI claims against the current code — deterministically, with zero egress (no API calls, nothing leaves the runner). Use it as a CI gate so a merge can't land while an advisor's cited file is fabricated or a checked claim no longer holds:

# .github/workflows/aegis-verify.yml
name: aegis verify
on: [pull_request]
jobs:
  verify:
    runs-on: ubuntu-latest
    steps:
      - uses: actions/checkout@v4
      - uses: actions/setup-node@v4
        with:
          node-version: 20
      - run: npx aegis-ai verify --ci
        # --ci exits non-zero when a claim is REFUTED or a cited file is FABRICATED
        # add --strict to also fail on unverifiable claims and unread citations

aegis inspect --ci works the same way for the deterministic governance checks.

CLI Commands

| Command | Description | |---------|-------------| | aegis init | Initialize .aegis/ governance structure | | aegis quickstart | Interactive setup with guided prompts | | aegis dashboard | Launch web dashboard | | aegis status | Show advisory summary | | aegis validate | Schema-check all .aegis/ files | | aegis history | List decisions with filters | | aegis inspect | Run deterministic governance checks | | aegis review | Review pending recommendations | | aegis brief | Generate AI executive brief | | aegis retro | Weekly decision intelligence report | | aegis session <role> | Generate advisor session prompt | | aegis focus add | Create standing directive | | aegis skills search | Search community skills | | aegis skills install | Install a community skill | | aegis adapt | Generate CLAUDE.md governance file | | aegis export | Generate STATUS.md for git | | aegis pause | Snapshot state for next session | | aegis resume | Pick up where you left off | | aegis vision | Start strategic vision cycle |

Dashboard

A newspaper-style, server-rendered dashboard. Primary tabs:

• Portfolio — Cross-project pulse: real-signal measurements (baseline-relative work patterns, 90-day velocity baseline, per-project trends — no pseudo-scientific scores).
• Overview — Editorial verdict on the board's standing (templated from advisor utility vectors, no LLM), six-advisor sigil grid with weights, recent decisions across the portfolio, sidebar with portfolio stats and alerts.
• Assessments — Multi-advisor assessments rendered as a numbered Live Disagreement thread. Each turn shows the advisor's sigil, stance (pro / con / neutral / concern), vote (1–5), and one-line summary; full assessment expands inline. Multi-round runs ("openings → rebuttals") stream turn-by-turn over WebSocket while LIVE mode is in flight. Header shows a Dissent Index (0–100, weighted spread of votes). Completed assessments get a deterministic Fact Check: refuted claims are flagged with the exact git/grep evidence.
• Logbook — Your coding logbook: weekly digest, streak, heatmap, and session feed across projects.

Under the More cluster: App Tour, Codebase Map, Skills.

Governance surfaces that don't have a dashboard tab (Advisors, Actions, Focus Areas) are accessible via the CLI and MCP server.

Skills tab

• Install community skills from skills.sh or GitHub URLs
• Compare any two local skills side-by-side
• Customize (merge edits into a new named skill, preserving provenance)
• Diff against upstream — check any skill against its source, see line-by-line changes, apply or keep local (with undo)
• Rate skills 👍/👎 — feedback drives plugin ranking during assessments

How It Works

All state lives in .aegis/ — human-readable YAML files, git-friendly, fully offline.

.aegis/
  config/ceo.yaml       — project governance config
  assessments/           — advisor assessment results
  decisions/             — decision records (by tier: T1-T4)
  directives/            — focus areas / standing orders
  proposals/             — pending approval items
  retros/                — retrospective snapshots
  skills/                — community skill plugins

Decision Tiers

• T1 (Autonomous) — Bug fixes, formatting, test additions
• T2 (Notify) — New features within existing commands
• T3 (Propose) — New commands, schema changes, new dependencies
• T4 (Escalate) — Breaking changes, security-sensitive, public API

MCP Integration

When connected via aegis-mcp, Claude Code gets governance tools that let it:

• Start sessions with project context (aegis_start)
• Validate changes against governance rules (aegis_validate)
• Log decisions with proper tiering (aegis_decide)
• Review advisor recommendations (aegis_review)
• Track decision outcomes (aegis_outcomes)

Desktop App

A native desktop app (Tauri) is available for macOS, Windows, and Linux. Download from GitHub Releases.

Privacy & Security

aegis is local-first: it runs on 127.0.0.1, sends no telemetry, and stores everything as plaintext under .aegis/ on your machine. The only outbound traffic is an anonymous, opt-out version check and — only if you choose live mode — your data to your own LLM provider. Full details:

• PRIVACY.md — the exact data-egress map and how to minimize sharing
• SECURITY.md — threat model, protections, and how to report a vulnerability

License

FSL-1.1-MIT — Functional Source License. Free to use, modify, and self-host. Converts to MIT after 2 years. The MCP server (aegis-mcp) is MIT licensed.