MCP server

Query geolocated world event signals (Delta Signals: geopolitical, security, disaster, and infrastructure events distilled from global news media, AI-enriched with severity/GEOINT scores and satellite-collection recommendations). Filter by bounding box, date window, and category. Costs 3 token(s) per call, charged to the API key owner's balance.

Result (structuredContent)

Aggregate statistics over the signal corpus — total event count plus per-category and per-day breakdown (trend) for a bounding box and date window. Cheaper than query_signals (returns roll-ups, not rows). NOTE the unit: `total` counts article-deduped events (meta.population = article_deduped_events), which is NOT cluster-collapsed, so it is >= the query_signals count for the same window. Costs 1 token(s) per call.

Result (structuredContent)

Geographic hotspots — signal density grid-binned into cells, ranked by event count, each with peak severity, the categories present, and up to 5 representative event_ids (trace a cell back to its signals). Use to find WHERE activity is concentrating. NOTE the unit: cells count satellite-observable points (meta.population = rs_observable_points); meta reports source_point_count and dropped_by_geo_count / dropped_by_severity_count so point_count is fully accountable. Costs 1 token(s) per call.

Result (structuredContent)

Fetch the Daily World Brief — an AI-synthesized OSINT/GEOINT digest of the previous UTC day's worldwide event signals (headline, executive summary, top developments with why-it-matters and what-to-watch, per-theme roll-up, ranked signals). Free of token charges. The result includes a freshness object (brief_date, age_hours, is_stale) — if is_stale is true this is the latest published brief and a newer day is not yet available, so relay it as possibly out of date.

Result (structuredContent)

Check the calling key's remaining token balance and plan capabilities — monthly allocation, tokens used this period, tokens remaining, and whether the plan includes AI tools over the API (assess_signal / ask_analyst). Use this to pre-flight a metered call: decide whether enough balance is left before spending. Free of token charges.

No arguments.

Result (structuredContent)

Search the satellite imagery catalog (Sentinel-1, Sentinel-2, NISAR L-band) for scenes over an area and date window — the natural follow-up to a signal (find imagery over the event location). Returns scene metadata (id, datetime, footprint, cloud cover, platform, orbit geometry, coverage, catalog link) — no imagery bytes. Pass eventDate to classify each scene timing=pre/post/same_day_unknown (a same-day scene is same_day_unknown, never post, without a real event time) and get pre/post bracketing + window_status + SAR sar_pair_status in meta. Pass eventPoint [lon,lat] and/or eventAoi [minLon,minLat,maxLon,maxLat] to get each scene's target_relation (covers_event_geometry = pure geometry gate; usable_for_analysis additionally requires acceptable cloud for optical, so a 99%-cloud scene is geometry-covering but not analysis-usable) — so a scene that only clips the wide bbox is not mistaken for covering the event. Costs 2 token(s) per call.

Result (structuredContent)

Plan the imagery evidence for ONE event in a single call, instead of guessing collections one at a time. Give the event_id and what you are trying to establish (damage_assessment | flood_mapping | wildfire_assessment) and the SERVER runs the deterministic plan for that goal: it always checks BOTH in-app sensors — sentinel-1-grd (SAR: all-weather, the only look that survives cloud and night) and sentinel-2-l2a (optical, human-legible) — exactly once each, against the event’s own footprint and a pre/post window around its date. The result carries, per search, why it was made, how many scenes came back, how many actually COVER the event and are usable (cloud-obscured optical does not count), the SAR pair status, and whether a pre/post bracket exists. It also states screening / detection / identification capability: Sentinel-1/2 screen and detect at facility scale, and only object-level identification needs commercial VHR — a VHR recommendation NEVER invalidates what the free catalog already showed. Prefer this over several search_imagery calls for the same event: it cannot miss the SAR look and cannot repeat a search. An event with no resolvable footprint is refused rather than planned against a guess. Costs 4 token(s) per call (it deliberately issues two catalog searches).

Result (structuredContent)

WHERE — and with what class (and therefore cost) of satellite — is observation most worthwhile right now? Crosses each event’s composite IMPORTANCE (severity × source breadth × market relevance) with the SPEC CLASS its required resolution demands: coarse (≤100 m), hr (≤10 m, Sentinel-class free data) or vhr (sub-metre, commercial tasking). Returns how many imageable events fall in each class with their mean importance, plus the top targets with importance, required class and AOI. Use it to triage a theatre before spending on imagery: a high-importance event that only needs hr is answerable with free Sentinel data, while a vhr one is what a paid order is for. Deterministic — no LLM, no per-event cost. Costs 1 token(s) per call.

Result (structuredContent)

Which events in a window can a given in-app sensor actually RESOLVE? Evaluates the FULL set (not just the top few) against each event’s precomputed required resolution, and returns how many are observable vs not, the breakdown by required resolution, and the top observable events with the ready-made imaging rationale. Observability here is resolvability — whether the physical mark is large enough for the sensor: Sentinel-2 is ~10 m optical (needs daylight and clear sky), Sentinel-1 is SAR (all-weather, day or night). Prefer this over asking about events one at a time: it is exhaustive AND cheap, and it is the honest way to answer "what can we actually see" before committing collection effort. The population is deliberately UNGATED by tasking readiness — it answers "what could this sensor resolve", not "what may we task" — so its total sits above rank_imaging_priority and counts a different unit from query_signals' clusters; see population_detail, and read collection_ready per event for taskability. Deterministic — no LLM. Costs 1 token(s) per call.

Result (structuredContent)

WHEN can this place next be imaged, and by WHAT — the timing half of collection planning. Propagates current orbital elements (SGP4 over day-cached CelesTrak two-line elements) for seven families and returns the access windows over a target: Sentinel-1 (SAR), Sentinel-2 and Landsat (free, SYSTEMATIC — routinely collected, so near-certain), plus WorldView, ICEYE, Capella and SkySat (commercial, AGILE — taskable ACCESS opportunities that require a paid order and are NOT guaranteed collects). Each pass carries acquisition/loss times, the closest-approach instant, peak elevation, OFF-NADIR angle, ground distance, ascending/descending, solar elevation and whether the target is sunlit (optical needs light; SAR does not), and the age of the element set it was computed from. Use it to answer "when is the next chance to see this", to choose between waiting for a free systematic pass and paying to task an agile one, and to time a pre/post change-detection pair. Give lat/lon, or a bbox whose centre is used. Horizon is capped at 7 days. If elements cannot be retrieved the result says so (retrieval_ok:false) — that means timing is UNAVAILABLE, never "no passes". Every pass is a GEOMETRIC access opportunity computed from orbital elements and swath width (geometry_only:true, acquisition_plan_verified:false): no operator collection plan is consulted, so this is when a sensor COULD see the target, never a confirmed acquisition schedule. With no start_date the window begins NOW, so the first pass listed is always still ahead. Costs 2 token(s) per call.

Result (structuredContent)

Run an AI RS (remote-sensing) deep-dive assessment for a specific signal: what to observe, recommended sensors, and a collection window. `eventId` is the `id` from query_signals. The result also carries a deterministic `context` block (event id/date, normalized target, AOI bbox, observability + quality verdict, and an `imagery_handoff` giving the exact bbox + event_date to pass to search_imagery for REAL pre/post scene candidates) — turning the assessment into an actionable collection plan. Costs 5 (quick) or 15 (deep) tokens, charged to the key owner's balance. A prior assessment for the same signal is cached (no re-charge). The exact charge and remaining balance are in the result meta.tokens. Signals that are not satellite-observable (observability:"not-observable" — e.g. political statements or broad-area events with no imageable physical mark) are rejected BEFORE any charge, so pre-filter with query_signals observability:"observable" to spend only where imagery helps.

Result (structuredContent)

Ask the Delta Analyst an OSINT/GEOINT question. Runs an agentic multi-step analysis over the signal corpus and returns a structured brief (summary, findings with collection recommendations, assessment, citations). Costs 5–123 tokens (usage-based, metered by the compute the question actually uses; charged ONCE, when the run completes; the exact charge and remaining balance are in the result meta.tokens). Durable async: the run is ENQUEUED and returns {status:"processing", job_id} immediately, then completes in a background worker — so it is never lost to a client timeout. Timing: most questions finish in ~30–90s; a complex brief (satellite-imagery lookups or many sources) can take 2–3 minutes. Fetch the finished brief by calling get_analyst_job with the job_id (poll every ~10–20s), or ask_analyst again with the SAME idempotencyKey (no second charge).

Result (structuredContent)

Fetch the status and result of an ask_analyst run by job_id. Job status is ONLY "processing" (still running — poll again in ~10-20s), "done" (with the finished brief and meta.tokens), or "error". When done, result_quality.status ("complete" | "partial", with any issues[]) reports whether the structured output is fully populated — this is SEPARATE from the job status (a done job can carry a partial result). Free of token charges — the run itself is charged once on completion. A job is visible only to the API key owner that created it.

Result (structuredContent)

Read the LEDGER of claims this key has been given — every factual assertion the Analyst made, with its evidence class (CONFIRMED / REPORTED / PARTY_CLAIM / ASSESSMENT), how many INDEPENDENT source families backed it, and the publishers. The point is the time axis: when a later answer restated the same assertion, the claim carries the link and says which way the evidence moved — restated_only surfaces those chains, and downgraded_only isolates the cases where a later answer was LESS sure than an earlier one, which is where this product contradicted itself. Use it to audit what you were told before acting on it, or to check whether an assertion has since weakened. Scoped to your own key; no other caller’s claims are visible. Free of token charges.

Result (structuredContent)

What actually CHANGED about the events in an area — not which articles are new. Polling query_signals and diffing the ids finds new REPORTING: a 2025 attack re-reported in 2026 with a fresh attribution is indistinguishable from a fresh attack. This returns changes as changes, each with what it was and what it became. Two kinds, labelled: `world` = the event's own state moved (a death toll revised, a perpetrator named, a report disputed); `measurement` = what WE can now see moved (a location resolved, post-event imagery arrived, a SAR before/after pair became ready). This is the right source for "what is new since last time" and for a standing watch. By default only changes worth a person's attention are returned — set notable_only=false for the full ledger, including changes deliberately judged too minor to notify on.

Result (structuredContent)

The full history of ONE event: its current state, and every change in the order it happened. Answers "is this a new event or an update to an old one" — the distinction a news feed cannot make. Returns the canonical event (where, when it HAPPENED as distinct from when it was reported, casualties, attribution and how that attribution is grounded), a timeline of developments, and every source article behind it. `occurred_at_basis` says whether the occurrence time was stated, absent, or never measured — those are three different things. `history_incomplete` marks an event that already existed when tracking began. Free of token charges: the discovery call that found the event was already metered.

Result (structuredContent)

Put an area under CONTINUOUS watch: save a question plus a bounding box and Delta re-answers it on a schedule, notifying only when the answer actually changed. Creating one is FREE. Each time it fires it runs ask_analyst and is metered like any Analyst question, so the cost is per CHANGE, not per check: a deterministic pass over the corpus decides whether anything new crossed the reporting bar, and quiet periods never invoke the model or charge anything. Returns projected_monthly_tokens_typical (the observed median cost per run × this cadence) and projected_monthly_tokens_max (the true ceiling: every check fires AND every run reaches the per-question cap), so the cost is visible before committing. Cadence and how many orders you may hold are set by your plan; the error says which limit you hit. Use it when the question is "tell me when this changes" rather than "what is happening right now".

Result (structuredContent)

List the layer sets saved on this account — the named map configurations a user builds in the app, each with how many layers it holds, its tags, when it was last changed and when it was last opened. Use it to answer "what have I saved" and to tell an idle configuration from a working one. The layer tree itself is not returned: it is an internal format, and opening it is the map’s job. Free of token charges.

Result (structuredContent)

Read one saved layer set by id: its name, description, layer count, size, tags and timestamps. The serialized layer tree is deliberately not exposed (contents:"not_exposed") — it is an internal representation, not a public contract. Free of token charges.

Result (structuredContent)

List the data this account uploaded to the map — name, format, size and when it was added — plus the formats the uploader actually accepts. Uploading itself happens in the app (it is a file transfer, not a JSON call), so use this to see what is already there and to answer format questions correctly: GeoJSON and GeoTIFF are accepted, a shapefile or KML is not. Free of token charges.

Result (structuredContent)

List the standing orders on this key, with each one’s cadence, watched area, when it last checked, when it last actually fired, and how many consecutive checks found nothing (quiet_checks — a high number means the watch is not earning its place). Also returns how many orders the plan allows and how many remain. Free of token charges.

No arguments.

Result (structuredContent)

Delete a standing order by id, or pause/resume it instead by passing active=false/true. Pausing keeps the order and its history; deleting removes both. Neither costs tokens. A paused order still counts against the plan limit, so delete rather than pause when you want the slot back.

Result (structuredContent)

List the places under continuous satellite measurement on this key (Delta Monitor), with each area’s metric, most recent value, change since the previous measurement, whether that value was flagged anomalous, and coverage — how many acquisitions were measured versus how many exist. Coverage window_total is null when the catalog total is UNKNOWN; null never means zero. Also returns how many areas the plan allows and how many remain. Free of token charges.

No arguments.

Result (structuredContent)

Fetch one monitored area with its full measurement history — every acquisition that was measured, its value, and whether it was flagged anomalous. This is the time series behind the number that list_monitored_areas reports, so use it to answer "is it going up", "when did it change", or "how unusual is today". Anomaly flags come from a median-absolute-deviation test on the series, not a fixed threshold. Free of token charges; returning the full history requires a plan that includes data export, so it can be refused with a forbidden error.

Result (structuredContent)

Put a place under continuous satellite measurement: pick an area and what to count, and every new Sentinel-1 / Sentinel-2 / VIIRS acquisition over it is measured automatically from then on. Use this when the question is about a quantity at a fixed place over time ("how many ships are alongside", "how much has burned", "is the water receding") rather than about events, which is create_standing_order. Creating is free; each automatic check costs a small number of tokens only when it finds new imagery.

Result (structuredContent)

Measure terrain height from the Copernicus DEM GLO-30 — a point (lat + lon), an area (bbox), or a drawn polygon, for which the statistics are computed over the samples INSIDE the ring rather than its bounding box. Returns min, max, mean, median, p10, p90 and relief (max − min), the number that governs SAR layover and shadow severity. Free of token charges. Three things the result carries that any answer must respect: it is a SURFACE model (buildings and tree canopy included, so not bare ground); `downsampled: true` means a large area was read below the 30 m posting, so min/max are smoothed inward and the relief is a floor rather than an exact figure; and `covered: false` or `tiles_missing > 0` means open ocean where the model has no data — which is absence, not 0 m. Heights are orthometric on the EGM2008 geoid, not ellipsoidal. Cite the returned `attribution` wherever a height is shown.

Result (structuredContent)

Compute FROM the terrain rather than reading heights out of it (that is lookup_elevation). Free of token charges; both operations come from one Copernicus DEM GLO-30 read. `operation: 'sar_geometry'` over a bbox, with an incidence angle and a look azimuth, returns the share of the area lost to LAYOVER and to SHADOW, the share merely foreshortened, and the mean LOCAL incidence angle — the arithmetic behind whether radar can use that ground. It changes with the pass direction (measured on one volcanic flank: 2.6% layover looking west against 0.7% looking east at the same 35°), so the geometry is required rather than assumed. `operation: 'profile'` between two points returns the ground along the line and a LINE-OF-SIGHT verdict including Earth curvature: whether the ends can see each other, where the terrain first rises above the sight line, and the worst clearance. Both results state the sample spacing they were computed at — a coarser grid reads flatter, and therefore more observable, than the ground is. A SURFACE model, so canopy and buildings are included; over a surface model they block a sight line as they would in reality. Cite the returned `attribution`.

Result (structuredContent)

Research one signal's location further and, if the sources genuinely narrow it, store and return the better coordinate. Most signals are geolocated once, for free, from a single article's place string: measured across production, about two thirds sit at ±5 km or have no bounded radius at all. This reads the source article and live search for a source that names something finer — a facility, bridge, district, port — resolves it, and verifies it before accepting it. **You are charged only if the precision actually improves**, between 3 and 29 tokens by what the run consumed; a signal that cannot be narrowed costs nothing, and a signal an earlier caller already refined is returned free. Read the result honestly: `improved: false` with a `reason` is the common and correct outcome when no source is more specific — it does NOT mean the event is unlocated, and it is not a failure to retry. When it does improve, `evidence` lists the sources that named the place; a location without evidence is never stored, so do not present a refined point without citing them. `footprint_kind` says whether the answer is a POINT or an AREA — if it is an area, describe an area, because an admin-level place has no finite radius and drawing it as a dot misstates where the event was. `uncertainty_m: null` means unbounded, not unknown.

Result (structuredContent)

Measure a spectral index over an area, scene by scene, back through the Sentinel-2 archive — the answer to "how has this changed since <year>". Give a polygon (or a bbox), an index (ndvi, evi, savi, ndmi, ndwi, mndwi, ndbi, ndsi, nbr, iron-oxide, clay, ferrous) and a date range; each scene is measured over the samples INSIDE the ring, and the result is the per-scene mean/median/min/max with a first-to-last change. Costs 0.5 tokens per scene actually measured; a scene that fails to read is reported in `skipped` and is not charged. **Call it once with `estimate_only: true` first** — that is free and returns how many scenes match, the real date span available, and what measuring them would cost, so the user can agree to the spend. Three things the result carries that any answer must respect: the Sentinel-2 archive begins 2015-06-27, so an earlier start is clamped and `clamped_to_archive` says so (those years are genuinely unavailable, not empty); at most 24 scenes are measured per call, so a longer period comes back as a SAMPLE and the rest appear in `skipped`; and `trend` compares the first and last measured scene only — it is not a fitted rate, so do not attach a slope or a confidence to it. `scenes_found` is what the catalog holds; `scenes_examined` is the newest page this call read, so `candidate_date_span` is the edge of that page and NOT how far the archive reaches — read `notes` before saying when coverage begins. For a SAR quantity, or to keep measuring as new imagery arrives, use create_monitored_area instead. Cite the returned `attribution`.

Result (structuredContent)

Count vessel-like targets in ONE SAR scene over an area, using CFAR detection on Sentinel-1. Costs 5 tokens per detection; a scene the worker refuses (an unsupported combination, an area with no usable sea) is not charged. The natural sequence is search_imagery over the area, pick a scene, then this with its `collection` and `item_id` — radar sees through cloud and at night, so the count works when optical would not. Read `caveats` before reporting the number: it states when the land mask was unavailable (shoreline structures may be counted as vessels), how far from the coast detections were excluded (the default excludes vessels alongside a quay), and when the scene covers only part of the requested area — a partial-coverage count must never be compared with a full one as though the difference were vessels. This measures one scene at one time. To track a berth or an anchorage over time, create_monitored_area with metric `ships` measures every new acquisition.

Result (structuredContent)