AI intelligence suite
What it does
The Intelligence Suite is a set of OpenClaw skills that run on fixed schedules. They scan RSS feeds, financial APIs, news sources, and public data, then send source-linked briefs to Telegram.
The system pushes relevant changes instead of relying on someone to check a dashboard.
The June 2026 research update added persistent memory, source provenance, retrieval checks, model routing, and explicit escalation thresholds. The operating rule is simple: collect evidence, retain only useful durable memory, retrieve it deliberately, and keep enough trace to explain why a signal was raised.
Active skills
Geopolitical and defence
| Skill | Schedule | Purpose |
|---|---|---|
| Conflict Dashboard | Hourly | Real-time Middle East SITREP, military actions, diplomatic developments, oil/Strait of Hormuz status, casualty tracking |
| Conflict RSS Scanner | 4x daily | Automated RSS ingestion from conflict news sources, severity scoring, SQLite storage |
| AI Defense Ethics Monitor | Daily | Tracks the ethical flashpoints at the intersection of AI and military/defense, Palantir, Pentagon, OpenAI contracts |
| Defense AI Intel | Daily | Broader defense-AI industry intelligence, procurement signals, policy shifts, capability announcements |
| Middle East SITREP | Daily 7 AM | Cross-referenced intelligence from Al Jazeera, Times of Israel, BBC, with X/Twitter verification |
Financial and portfolio
| Skill | Schedule | Purpose |
|---|---|---|
| Flash Crash Guardian | Every 5 min | Monitors portfolio for sudden >5% drawdowns; autonomous position flattening capability |
| Portfolio Tracker | Bi-hourly (market hours) | Live pricing across ASX, US equities, and ETFs, validates data freshness, flags >3% movers |
| Nexus Daily Recommendations | Daily 9 AM | Integrated NEXUS engine status + all share positions with P&L, sector exposure, and conflict cross-reference |
| Earnings Reports | Daily | Upcoming earnings calendar for held positions, alerts before market-moving announcements |
| Cost Tracker | Daily | API spend and infrastructure cost monitoring |
Technology and AI leadership
| Skill | Schedule | Purpose |
|---|---|---|
| Tech Leaders Daily Intel | Daily 9:15 AM | Scans web for latest from tracked tech leaders, Dario Amodei, Jensen Huang, Satya Nadella, and others. Produces project proposals from signals |
| Anthropic Legal Tracker | Daily | Monitors Anthropic's legal exposure, supply chain risks, and the Pentagon standoff |
| GTC Digest | Event-driven | Categorises and summarises NVIDIA GTC announcements |
| Meta Model Intel | On-demand | Structured intelligence on Meta's AI model pipeline, Avocado AI, and competitive positioning |
| M5 Mac Watch | 2x weekly | Monitors Apple M5 Mac release announcements, only alerts on confirmed launches |
Home and operations
| Skill | Schedule | Purpose |
|---|---|---|
| AI Vision Sentry | Hourly | Computer vision sweep of all home cameras, person/vehicle/anomaly detection |
| Home Optimus | Hourly | Smart home optimisation scoring, battery levels, energy efficiency, device health |
| Doge Home Energy | 2x daily | Home energy and efficiency reporting with HA sensor data |
| Security 360 | 4x daily | Comprehensive security assessment, network, devices, access patterns |
| Network Watchdog | Every 6h | Multi-path connectivity testing with diagnostic reasoning |
Architecture
Each skill follows the same flow:
- Collect data from RSS feeds, APIs, web pages, and Home Assistant sensors.
- Score severity, remove duplicates, and check freshness.
- Compare sources and identify material changes.
- Send the result to Telegram with timestamps and citations.
- Store state in SQLite, JSON files, and memory logs.
- Apply retention, provenance, supersession, and stale-signal rules.
- Record the source list, model route, confidence, and escalation reason for high-impact briefs.
Every report includes a Data as of: <timestamp> header. The report flags stale data explicitly.
Model routing
The router selects a model by task:
- Claude Sonnet 4.6 handles most synthesis, analysis, and reporting.
- Claude Opus 4.6 handles complex multi-source reasoning.
- Local Ollama models handle lightweight classification.
Routing also considers service health. If the primary model is degraded, the suite uses a stable fallback rather than stall the cycle. A timely, source-grounded summary with clear uncertainty is more useful than a late report from a stronger model.
Memory and evaluation
Persistent memory is useful only when it is selective. The suite retains durable memory for recurring entities, thresholds, known false positives, prior escalations, and user decisions. It does not retain every article or every model-generated paragraph. That distinction follows the June 2026 Agent Memory research: construction cost, retrieval cost, and stale context risk are operational design constraints.
Evaluation checks whether the agent retrieved the right prior context, preserved source links and timestamps, and escalated while the signal was still actionable. For example, a new Anthropic legal article should be compared with earlier supply chain, copyright, and government contract signals.
Delivery channels
All intelligence is delivered via Telegram with topic-based routing:
- Security: network audits, Security 360, and camera alerts.
- Markets: portfolio updates, flash crash alerts, and earnings.
- General: daily briefs, technology leader updates, and conflict reports.
Design choices
Timing changes the value of information. A conflict update read six hours late may no longer support a useful decision. A portfolio alert at 3% gives the operator more time to assess a move than an alert at 15%.
Each skill sends a bounded update with the source, timestamp, and context needed to assess it. Routine noise stays out of the channel.
A pattern worth considering for a non-technical stakeholder version of this suite: Claude Cowork shares the same agentic core as Claude Code but needs no terminal, so a portfolio manager or compliance reviewer could ask it to pull the latest signal from a connected channel and summarise the last week's high-relevance items, rather than reading raw briefs. This suite's own routing and evaluation model — bounded updates, source and timestamp attached, routine noise filtered before delivery — is exactly the discipline a Cowork-style deployment would need to inherit to stay trustworthy.
References
- OpenClaw Architecture: Living Architecture
- AI Research 2026: Reasoning, Agents, Memory, and Production Inference
- Anthropic Legal Tracker: Skill Documentation
- AI Defense Ethics: Skill Documentation
- GTC Digest: Skill Documentation