Gdelt Mcp Server

Overview

News and TV coverage analysis from the GDELT Project: global news in 65+ languages, and historical TV transcripts across 150+ stations, with coverage dates varying by station. Article search and relative timespan windows cover the last 3 months; explicit dates reach older windows in the coverage timeline's volume and tone modes, the language and country breakdowns, and the tone distribution. Search articles and clips, track coverage spikes, analyze tone, and compare coverage across languages, countries, stations, and shows.

Tools

Capability reference

gdelt_search_articles _tool

• Full GDELT query syntax: phrases, boolean OR, exclusion, filter operators (sourcecountry:, sourcelang:, domain:, theme:, tone</tone>), proximity (near20:) and repetition (repeat3:) — gdelt_search_themes finds the identifiers theme: takes
• Configurable sort (relevance, dateDesc, dateAsc, toneDesc, toneAsc, hybridRel) and fetch count, up to 250 per call; larger maxRecords values clamp to 250 by default
• Returns URL, title, publication date, domain, language, source country, and social image URL
• Returns every fetched in-window article. At the provider ceiling, continuationWindows gives narrower date windows to re-query (de-duplicate by url). The ceiling is not a cursor; an indivisible window can still contain more matches than GDELT returns
• Articles outside an explicit startDatetime/endDatetime window are dropped, and the notice counts them

gdelt_get_coverage_timeline _tool

• Four modes: volume (normalized % per timestep), volume_raw (matching-article counts with the monitored-article denominator norm), volume_with_articles (volume plus supplied top articles), and tone (average sentiment)
• Reference-bearing timesteps carry their first 3 links beside referenceCount, the number of references GDELT supplied, not all matching articles. points: ["<date>"] selects named timesteps with every supplied link; GDELT's reference list itself is capped at ten
• Pages whole timesteps with offset and limit (default 10, cap 400); nextOffset retrieves the next page. Each series' pointCount and peak cover the full series on every page
• Smoothing accepts 0–30 timesteps. Width N > 1 returns trailing moving averages: the first N−1 source steps cannot produce a value, and peaks can appear later. In volume_raw, smoothing changes value, not norm
• Time range uses timespan (up to 3 months) or explicit startDatetime/endDatetime; date resolution (15min/hour/day) is inferred from returned intervals. In volume and tone modes an explicit window may predate the last 3 months; volume_raw and volume_with_articles cover only the last 3 months
• A points date matching no timestep, or an offset past the last timestep, is rejected with the valid values

gdelt_get_tone_distribution _tool

• Histogram bins from approximately −30 to +30; each bin's count counts matching articles, while referenceCount counts supplied links. The default preview shows 3 links; bins: [<bin>] selects named bins with every supplied link. An unknown bin is rejected with the available bins
• Summary fields over the whole histogram: peakNegativeBin, peakPositiveBin, neutralPct (% of articles in the −2 to +2 range)
• Pages whole bins with offset and limit (default 10, cap 400), returning nextOffset when more remain. This is a snapshot across matching articles, unlike the tone timeline
• Time range uses timespan (up to 3 months) or explicit startDatetime/endDatetime, and an explicit window may predate the last 3 months

gdelt_get_coverage_breakdown _tool

• Breaks down by language or country into a multi-series time series: the top 10 series by summed value, with every other series named in otherSeriesLabels
• series: ["<label>"] returns only those series, complete, under selectedSeries; a label matching no series is rejected with the available labels
• Pages whole series with offset and limit (default 10, cap 75), returning nextOffset when more remain. Copy exact labels from the response for series selections
• Smoothing accepts 0–30 timesteps with the same moving-average disclosure as coverage timelines
• Time range uses timespan (up to 3 months) or explicit startDatetime/endDatetime, and an explicit window may predate the last 3 months
• Values are normalized shares of each source's media output, not article counts, so don't sum them across sources

gdelt_search_themes _tool

• Searches the identifiers in GDELT's GKG theme lookup (59,315 themes); each match is the identifier, its listed count, and a paste-ready operator such as theme:TAX_DISEASE_OUTBREAK
• Case-insensitive, with no stemming or synonyms: words match parts of an identifier (cyberattack → CYBER_ATTACK, plant disease → TAX_PLANTDISEASE). If nothing matches, it retries once with a trailing s dropped (protests → PROTEST) and the notice says so
• Ranked by exact match, then listed count (a static prevalence figure, not a live article total), then identifier
• Pages with offset / limit (default 25, cap 100); larger limits clamp by default. nextOffset appears only when more matches remain, and an offset past a non-empty result fails offset_out_of_range
• Does not queue behind the DOC and TV rate limit

gdelt_search_tv _tool

• Structured stations (e.g. ["CNN", "FOXNEWS"]) preserve the entire requested list. TV queries can also use station, network, or market selectors; verify IDs and coverage dates through the station catalog
• normalize toggles relative airtime % (default) vs. raw matching 15-second clip counts; optional dateres aggregation (hour/day/week/month/year). TV operators: station:, network:, market:, show:, context:
• Each series carries station (the ID the stations input takes, FOXNEWS) and stationName (GDELT's label, FOX News). If an ID can't be resolved, the series keeps stationName without station and the notice says so
• Pages points in date-then-station-name order (default 500, cap 20,000); pass nextOffset with the same inputs for the next page. The text surface's per-station count, sum, and peak cover only the returned page
• combined: true pools stations into a Combined timeline. Normalized values use pooled matching clips divided by pooled monitored clips, not an average of station percentages; no station ID is assigned to the combined series
• Smoothing accepts 0–30 timesteps and discloses leading-window loss and possible lag. Recognized upstream date resolution takes precedence over requested or inferred resolution
• Check station date ranges with gdelt_list_tv_stations before querying recent events

gdelt_get_tv_clips _tool

• Up to 3,000 clips fetched per call, sorted by relevance, date descending, or date ascending; larger maxRecords values clamp to 3,000 by default
• Each clip: show name, station and stationName (resolved as in gdelt_search_tv, so a series' station passes straight into stations), air timestamp, 15-second transcript excerpt, direct Archive.org link, and optional thumbnail
• The server requests whole clock hours, so any window width works; clips outside an explicit startDatetime/endDatetime window are dropped and counted in the notice (a timespan call drops nothing)
• Returns every fetched in-window clip. At the provider ceiling, continuationWindows gives narrower windows to re-query (de-duplicate by archiveUrl); whole-hour retrieval can still prevent reaching every match in a dense hour

gdelt_get_tv_breakdown _tool

• Station rankings report percentages of station airtime by default, or matching 15-second clip counts with normalize: false; resolvable rows include a station ID
• Show rankings always report percentages of matching clips, including when normalize: false. Labels are preserved exactly, but GDELT sometimes abbreviates them with ellipses, so they are not all reusable show: selectors
• Rows sort by descending value, then label, with offset / limit paging (default 25, cap 10,000), totalRows, and nextOffset

gdelt_get_tv_context _tool

• Returns the most frequent non-stopword terms co-occurring with the query
• Terms and scores come from GDELT's sample of the 200 most relevant matching clips, not all coverage; every non-empty response says so in notice. Compare windows or stations with a separate, narrower call for each
• Relative frequency scores 0–100, where the query term itself scores 100
• Terms sort by score and page with offset / limit (default 25, cap 25,000). totalTerms counts the returned vocabulary across all pages, not matching clips; nextOffset and truncated disclose remaining terms

gdelt_list_tv_stations _tool

• Returns every station with market, network, monitoring start date, and end date — or only those matching the optional stations, network, and market filters
• Filters match whole values case-insensitively and combine with AND (network: "FOX" does not match FOXNEWS); activeCount and totalCount count the returned stations
• A filter or station ID that matches nothing is named in the notice; isActive is true when the end date is within the last 24 hours
• Use to verify a station was active during a target period, or to find valid station IDs for the stations parameter on other TV tools

Features

Built on @cyanheads/mcp-ts-core: stdio and Streamable HTTP transports, pluggable auth (none / jwt / oauth), swappable storage (in-memory, filesystem, Supabase, Cloudflare KV/R2/D1), structured logging with optional OpenTelemetry tracing.

GDELT-specific:

• Shared request pacer across the DOC and TV tools: one request in flight at 1 req/5s, with a backoff (5s, doubling to 60s) after any rate-limit response
• In-process response cache for DOC and TV requests, bounded to 128 MiB. An identical request is answered from memory for 15 minutes (24 hours once its explicit window ended more than a day earlier), without a pacer slot, and concurrent identical calls share one fetch. When GDELT rate-limits the refetch, is unavailable (5xx, timeouts), or answers it empty, the stored answer stands in for up to 24 hours after its fetch (a timespan answer only while younger than the timespan), and the response says so in staleResponse; an invalid query, an unparseable body, or a cancelled call never falls back. A call holding such an answer waits on GDELT at most 45 s, and not at all while the pacer is cooling down after a rate limit
• Optional local GKG mirror that answers the four DOC tools when the DOC API refuses or returns nothing; see Local GKG mirror
• Structured stations arrays instead of raw TV query syntax
• GKG theme lookup held in memory after first download

Agent-friendly output:

• Query echo on every response: the original query and applied timespan, so calls chain without re-deriving parameters
• Typed series label fields ("Volume Intensity", "Average Tone", language/country names) instead of positional arrays
• Station metadata with an isActive boolean and ISO 8601 dates
• Zero matches return the normal output shape with empty arrays and a notice on how to broaden the query. On the DOC tools the notice also says GDELT returns empty answers when overloaded, so retry before treating the result as no coverage
• Provenance fields: staleResponse (fetchedAt, ageSeconds, reason) on every DOC and TV tool when a stored answer was served; source, attempts, and coverage on the four DOC tools while a mirror is configured

Getting started

Public Hosted Instance

A public instance is available at https://gdelt.caseyjhand.com/mcp — no installation required. Point any MCP client at it via Streamable HTTP:

{  "mcpServers": {    "gdelt-mcp-server": {      "type": "streamable-http",      "url": "https://gdelt.caseyjhand.com/mcp"    }  }}

Self-Hosted / Local

Add the following to your MCP client configuration file.

{  "mcpServers": {    "gdelt-mcp-server": {      "type": "stdio",      "command": "bunx",      "args": ["@cyanheads/gdelt-mcp-server@latest"],      "env": {        "MCP_TRANSPORT_TYPE": "stdio",        "MCP_LOG_LEVEL": "info"      }    }  }}

Or with npx (no Bun required):

{  "mcpServers": {    "gdelt-mcp-server": {      "type": "stdio",      "command": "npx",      "args": ["-y", "@cyanheads/gdelt-mcp-server@latest"],      "env": {        "MCP_TRANSPORT_TYPE": "stdio",        "MCP_LOG_LEVEL": "info"      }    }  }}

Or with Docker:

{  "mcpServers": {    "gdelt-mcp-server": {      "type": "stdio",      "command": "docker",      "args": [        "run", "-i", "--rm",        "-e", "MCP_TRANSPORT_TYPE=stdio",        "ghcr.io/cyanheads/gdelt-mcp-server:latest"      ]    }  }}

For Streamable HTTP, set the transport and start the server:

MCP_TRANSPORT_TYPE=http MCP_HTTP_PORT=3010 bun run start:http# Server listens at http://localhost:3010/mcp

Prerequisites

• Bun v1.4.0 or higher (or Node.js v24+).
• No API key required — GDELT is a free public API.

Installation

1. Clone the repository:

git clone https://github.com/cyanheads/gdelt-mcp-server.git

2. Navigate into the directory:

cd gdelt-mcp-server

3. Install dependencies:

bun install

4. Configure environment:

cp .env.example .env# edit .env if you need to override defaults

Configuration

All configuration is validated at startup via Zod schemas in src/config/server-config.ts. Key environment variables:

See .env.example for the full list of optional overrides.

Running the server

Local development

• Build and run:

  sh

  # One-time buildbun run rebuild
  # Run the built serverbun run start:stdio# orbun run start:http

• Run checks and tests:

  sh

  bun run devcheck   # Lint, format, typecheck, security auditbun run test       # Vitest test suitebun run lint:mcp   # Validate MCP definitions against spec

Docker

docker build -t gdelt-mcp-server .docker run --rm -p 3010:3010 gdelt-mcp-server

The Dockerfile defaults to HTTP transport, stateless session mode, and logs to /var/log/gdelt-mcp-server. OpenTelemetry peer dependencies are installed by default — build with --build-arg OTEL_ENABLED=false to omit them.

Local GKG mirror

The DOC API sometimes refuses requests for long stretches, or answers empty while overloaded. The optional mirror keeps a local SQLite copy of GDELT's Global Knowledge Graph (GKG), the English and translingual feeds GDELT publishes every 15 minutes on data.gdeltproject.org, so the four DOC tools (gdelt_search_articles, gdelt_get_coverage_timeline, gdelt_get_tone_distribution, gdelt_get_coverage_breakdown) can still answer. It is off unless GDELT_MIRROR_PATH is set. TV tools and theme search never use it.

When it answers. DOC is asked first (a stored answer to the identical request counts as DOC's, and wins over the mirror), and the mirror answers when DOC is rate-limited, unavailable, or returns nothing. An empty DOC answer is replaced only if the mirror finds at least one match. Invalid queries and cancelled calls never reach it. A mirror answer carries source: "gkg_mirror", attempts (what each source did), and coverage when the mirror holds less than the requested window, and the notice says why it answered. When the mirror cannot answer (no completed init, a query operator it can't evaluate, a window outside what it holds), DOC's error is returned with the mirror's reason in the message and a hint.

How its answers differ from DOC's.

• Keywords match titles and entity names (persons, organizations, locations), not article text. Translated articles keep their source-language titles.
• theme: matches the exact theme ID. sourcecountry: and country breakdowns use GDELT's 2015 domain-to-country file, which covered 83–92% of sampled rows; matching articles from outlets it doesn't list are left out of country series and counted in the notice.
• Counts, shares, and the volume_raw denominator are over mirrored articles, so they differ from DOC's figures.
• nearN:, repeatN:, the image operators (imagetag: and the rest), unknown name: operators, and comparisons other than tone/toneabs are not compiled; a sourcelang: or sourcecountry: value it can't resolve is refused too. It also refuses an unquoted keyword that is one character or contains a hyphen, an OR group that nests or contains an exclusion, an excluded OR group, and a query of more than 100 elements or 1,000 characters. The mirror declines a query that uses any of these, and the error or the attempts detail names which.

Enable it.

1. Set GDELT_MIRROR_PATH to a database file, and optionally GDELT_MIRROR_WINDOW_DAYS.
2. Build it once with bun run mirror:init from a checkout. It downloads every 15-minute slot of both feeds in the window, 2 s apart, resumes where it stopped if interrupted, and never runs on server startup.
3. Start the server. Bun uses its built-in SQLite; on Node, install better-sqlite3 too, or startup fails when GDELT_MIRROR_PATH is set.

With Docker, point the path at a file on a mounted volume (the image creates /var/lib/gdelt-mirror, owned by the bun user) and run the scripts in the container:

docker run -d --name gdelt -p 3010:3010 \  -v gdelt-mirror:/var/lib/gdelt-mirror \  -e GDELT_MIRROR_PATH=/var/lib/gdelt-mirror/gkg.db \  gdelt-mcp-serverdocker exec gdelt bun run mirror:init

bun run mirror:verify (or docker exec gdelt bun run mirror:verify) reports the SQLite integrity check, sync status, row count, and each feed's covered slots and gaps.

Disk and bandwidth. A stored article takes about 1.8 KB, plus about 3 MB for the domain-to-country table. A 7-day window (the default) runs about 3 GB in a typical week and up to about 5.6 GB at weekday peak volume; a busy 15-minute slot holds up to about 4,760 rows across both feeds. The first mirror:init for 7 days fetches about 1,344 files, up to about 12.5 GB, and takes 50–60 minutes, most of it the pause between files. Steady-state refresh downloads 0.56–1.71 GB a day. Size scales with GDELT_MIRROR_WINDOW_DAYS.

Refresh. On HTTP transport the server refreshes the mirror every 15 minutes once a first init has completed. On stdio, run bun run mirror:refresh from cron every 15 minutes. Don't run cron refresh against a store an HTTP server is also refreshing: the result stays consistent, but the downloads are duplicated.

Project structure

Development guide

See CLAUDE.md for development guidelines and architectural rules. The short version:

• Handlers throw, framework catches — no try/catch in tool logic
• Use ctx.log for request-scoped logging, ctx.state for tenant-scoped storage
• Register new tools via the barrels in src/mcp-server/tools/definitions/index.ts
• Wrap GDELT API calls: validate raw JSON → normalize to domain type → return output schema; never fabricate missing fields

Contributing

Issues are welcome. Run checks and tests before submitting:

bun run devcheckbun run test

License

Apache-2.0 — see LICENSE for details.