Build a Daily Situation-Briefing Agent with MCP Tools and Area Watch

Most people wire an AI agent to live data and then ask it one-off questions. The higher-value pattern is a routine: the same briefing, produced every morning, plus a standing watch on the places you care about. This post shows how to build that with the Off-Nadir Delta MCP server — writing the routine once as an instruction and letting the agent call the tools itself.

It is the operational third part of a short series: giving an AI agent live world-event awareness covers the surfaces, and from a world event to the satellite scenes over it covers the imagery chain. Here we make it recurring.

What is a situation-briefing agent, and what does it do each day?

A situation-briefing agent is an AI agent that runs a fixed intelligence routine on a schedule rather than answering a single prompt. Each day it reads the current world brief, lists the most severe recent events, and drills into the most significant one; for a watched area it can fetch only what has been updated since its last run. The value is consistency: the same structured read, every day, grounded in sourced events instead of a model's training snapshot.

Which tools does the routine use?

The Model Context Protocol lets a server expose tools (actions the agent takes) and resources (context it reads). A briefing routine needs only a handful of Off-Nadir Delta's tools; the MCP docs list the full, current roster.

Tool	Role in the routine
get_world_brief	The AI-synthesized World Brief for the previous UTC day — free
query_signals	Geolocated events for an area and window, filterable by severity band, with a differential mode (updated_since)
query_developments	What actually changed about the events in an area — new events and updates to existing ones, told apart
assess_signal	A remote-sensing assessment of one event: whether and how a satellite could confirm it
get_usage	Your remaining token balance — free

How do you run a daily world briefing with the MCP server?

Give the agent the routine as a standing instruction. It reads get_world_brief — the AI-synthesized World Brief (headline, executive summary, top developments, a per-theme roll-up and ranked signals), which costs no tokens — then calls query_signals with the top severity band (min_severity_band: 9) to list the day's most severe events, and assess_signal on the top one. You get a consistent morning read: what happened, what matters most, and what to look at next. In an MCP client this is a single instruction, for example:

Using the Off-Nadir Delta tools: read the latest World Brief, list the signals from the last
24 hours in the highest severity band, then run assess_signal on the first one.

Because get_world_brief and get_usage are free of token charges, the "read the brief and check my balance" part of the routine costs nothing; only the query_signals and assess_signal steps spend from your token balance. The severity-band filter needs a plan with event filters; on Free it is not applied, and meta.filter_clamp says so. The same brief is also published in the app as the Daily World Brief, so the agent's briefing matches what a human analyst would read in the product.

How do you watch a specific area of interest for change?

Call query_signals scoped to a bounding box. To make it a watch rather than a re-read, the signals tool supports a differential mode: pass updated_since and it returns only events updated after that timestamp — so a scheduled agent fetches the delta since its last run instead of the whole window. Results come back up to 500 signals per page, cursor-paginated, each with coordinates, a severity score, source links, and a recommended sensor. Over the REST API the same call is a plain request:

# Only armed-conflict events in a bounding box that changed since the last poll
curl -H "Authorization: Bearer ond_..." \
  "https://offnadir-delta.com/api/v1/signals?bbox=30.0,50.0,31.0,51.0&categories=armed_conflict&updated_since=2026-09-27T06:00:00Z"

When the question is "what changed" rather than "what was reported", use query_developments instead: it separates new events from updates to events already reported, and says for each change what the value was and what it became. An agent running this on a schedule turns it into a standing brief: "here is what changed in your area since yesterday, ranked by severity, and here is why it matters." If you would rather not run the schedule yourself, a standing order (POST /api/v1/standing-orders, or client.standing_orders.create in the Python SDK; over MCP, create_watch saves the area to your Watchlist) puts the area under continuous watch on the server — the same as the app's "Watch this area": it is checked weekly and notifies only when new events cross the reporting bar (subject to a per-plan limit on how many can run). For the definitions behind the feed — OSINT, GEOINT, and tip-and-cue — see the glossary.

How do you keep the agent's briefing grounded and honest?

Two habits. First, keep every claim traceable: each signal in Delta Signals links back to its original sources and to a location whose precision is stated, so the agent can cite evidence instead of asserting it — which is what makes the feed evidence-based rather than a black box. Second, preserve uncertainty: the assess_signal and ask_analyst tools return findings with calibrated confidence written to professional analytic standards (see what a daily intelligence report is), so the agent reports "moderate confidence, single source" honestly rather than overclaiming. In practice a well-behaved briefing agent states what it knows, what it infers, and what it cannot yet confirm.

What does a daily briefing agent cost to run?

The routine's read steps are free, and the analysis steps are metered on your token balance — there is no separate seat license or per-image contract. get_world_brief and get_usage cost no tokens; query_signals, query_developments, and search_imagery are billed per query; and assess_signal and ask_analyst are billed per call. Have the agent read the usage://current resource (or call get_usage) at the start of a run to pre-flight what it can spend, and it will degrade gracefully when the balance is low — brief first, deeper assessments only if budget remains. Both the API and MCP server are available on every plan, including Free; the only gate is your token balance. Current allocations live on the pricing page.

How do I connect the server to start?

Issue an API key from Account → Developer API, then register the server as an HTTP transport with your key as a bearer header. In Claude Code:

claude mcp add --transport http off-nadir-delta \
  https://offnadir-delta.com/api/v1/mcp \
  --header "Authorization: Bearer ond_..."

Any MCP client that supports HTTP transports connects the same way — Claude Code, Claude Desktop, Cursor, or VS Code — by pointing at https://offnadir-delta.com/api/v1/mcp and passing the Authorization header. Full client setup, the tool list, and the resource catalog are in the MCP server documentation.

When should you use an agent routine vs a REST cron job?

Use an agent routine over MCP when a person (or an assistant) should run it conversationally and read the result — a morning briefing you skim, an area watch you ask follow-up questions about. Use a REST cron job when the output feeds a machine — a dashboard tile, an alerting rule, a report generator — where you want deterministic JSON on a fixed schedule. They share the same token wallet and the same event layer, so the two combine: a scheduled REST poll can fill a common operating picture while an MCP-connected agent answers "so what changed, and why does it matter?" For the full REST surface and config blocks, see the event intelligence API and the API & MCP reference.