Conversational Dashboard
Governance operations through natural language. Ask questions, review escalations, explore decisions, run backtests — all in a chat interface.
Chat, Not Dashboards
Traditional governance dashboards bury critical signals under page navigation. GaaS takes a different approach: you talk to your governance system. Ask "what happened overnight?" and get a summary with data. Say "review that top escalation" and get the context package with approve/deny options. No clicks through menus. No tab switching.
The dashboard is powered by Claude, with 33 tools that map to the full GaaS API. Claude fetches live data, generates visual artifacts inline, and takes actions on your behalf — all within the conversation.
Architecture
You
Type a question or click a suggested action. Natural language, no syntax required.
Chat UI
React frontend with markdown rendering, live artifact preview, and suggested action chips. Streams responses in real time via server-sent events.
BFF (Express + Claude)
Backend-for-frontend server. Sends your message to Claude with 33 governance tools. Claude decides which tools to call, executes them against the GaaS API (up to 10 rounds), and streams text and artifacts back.
GaaS API
The governance engine. All data — decisions, escalations, learning, membranes — comes from the same API your agents use.
What You Can Do
The dashboard exposes five categories of governance operations through natural language.
Operational Intelligence 10 tools
- Command center — real-time metrics: decisions today, blocks, pending escalations, risk index, block rate, latency
- Attention queue — items needing immediate human attention, ranked by urgency
- Daily digest — yesterday's governance summary with block/modify/escalate breakdown
- Weekly summary — executive overview with week-over-week comparisons and top risks
- Evidence packages — generate compliance packages for auditors with sampled audit records and hash-verified chains
- Decision stream — filterable feed of all governance decisions (by verdict, risk level, agent, action type, pipeline mode, time range)
- Agent explorer — overview of all registered agents with risk patterns and compliance metrics
- Notification rules — configure alerts for escalation timeouts, risk thresholds, and anomalies
Escalation Management 8 tools
- List and filter escalations by status, reviewer, risk score, trigger source
- View full escalation context: intent summary, policy evaluation, deliberation positions, audit reference
- Submit review decisions: approve, modify, or deny with notes and conditions
- Reassign escalations to different reviewers
- Cancel escalations when the original intent is no longer relevant
- View escalation statistics: resolution rates, average review time, timeout rates
Learning & Calibration 9 tools
- View current calibration: agent weights, deliberation threshold, risk dimension weights, circuit breaker status
- Browse calibration history: parameter changes, triggers, impact assessments
- Explore insights: policy hotspots, override patterns, risk drift, latency bottlenecks
- Run backtests: simulate how a parameter change would have affected historical decisions
- Submit feedback on decisions: flag false positives, false negatives, or confirm good catches
- Report governance incidents linked to specific audit records
- Check Collective Knowledge Base updates from the GaaS network
Onboarding & Membranes 2 tools
- View membrane status: draft, shadow, or live with full UCM and policy set
- Check shadow mode statistics: verdict distribution, average risk, policy coverage
Audit & Decisions 2 tools
- Retrieve any governance decision with full reasoning chain, risk assessment, and blocking policies
- Access complete audit trails covering all 5 pipeline stages
Visual Artifacts
When data needs to be seen rather than described, Claude generates live React components inline in the conversation. Charts, tables, metric cards, and status indicators render directly below the response text.
What gets rendered
- Bar, line, area, and pie charts (decision trends, risk distributions, verdict breakdowns)
- Data tables with clickable rows (escalation queues, decision streams, agent lists)
- Metric cards with trend indicators (decisions today, block rate, risk index)
- Status indicators (pipeline health, circuit breaker state)
- Interactive buttons that trigger follow-up messages (approve/deny/review actions within artifacts)
Artifacts are expandable to full-screen and include a code toggle for transparency. Every visualization is generated from live GaaS data — nothing is mocked or pre-built.
Suggested Actions
After responses with clear next steps, clickable action chips appear below the message. These guide workflows without requiring you to type.
Example Conversation Flows
Morning check-in
Escalation review
Calibration backtest
Setup
| Variable | Required | Description |
|---|---|---|
ANTHROPIC_API_KEY |
Yes | Claude API key for the BFF conversation engine |
GAAS_API_URL |
No | GaaS API server URL (default: http://localhost:8000) |
GAAS_API_KEY |
No | API key for the GaaS API (omit if running in open mode) |
cd apps/dashboard
npm install
npm run dev
This starts both the BFF server (port 3001) and the Vite dev server concurrently. The frontend proxies /api requests to the BFF automatically.
Related Pages
- Getting Started — onboard, integrate, and go live
- Shadow Mode — run the full pipeline without enforcement
- SDKs — Python, TypeScript, and Java client libraries