GPU-Accelerated Geospatial Compute for the Agent Era
As of August 2026, the only commercially-supported GPU-accelerated geospatial MCP server we have been able to identify. Distributed as pure-Java Maven artifacts — no first-party native code, no GDAL native bindings, no separate native install step. Nearly every operation is callable by an AI agent through a native Model Context Protocol server.
Three short answers for the technical evaluator.
ForgeGIS is a GPU-accelerated geospatial compute library and a native Model Context Protocol server, both written in Java. It runs raster, hydrology, visibility, spectral, ML, and point-cloud workloads on the GPU through Java-side compute kernels, alongside a large vector and topology surface in pure Java — and hands almost its entire catalog to AI agents through MCP.
Serious geospatial work has historically meant assembling Python, C++, and native libraries — GDAL, JTS, PostGIS, CUDA — and accepting either CPU-bound throughput or a build pipeline tied to native dependencies. ForgeGIS removes both costs: GPU performance without leaving the JVM, and agent-callable surface without an integration layer.
Pure-Java Maven distribution. No first-party native code, no GDAL native bindings, no separate native install step. GPU dispatch is via the JOCL OpenCL bindings (JNI bridge bundled in the JOCL JAR, extracted at runtime); the system OpenCL runtime is standard system software bundled with GPU drivers. Deployment is one JAR.
One library, two surfaces — one designed for agent composition, one for typed single calls.
The MCP server (forgegis-mcp) exposes ForgeGIS through two complementary surfaces, so an
AI agent can either compose a multi-stage workflow in a single call or invoke a specific operation
directly with typed inputs.
Composable JSON pipeline format. Agents assemble multi-stage workflows (read → reproject → mask → classify → write) inside a single MCP call, including branching DAGs — the same document format Studio saves.
Strongly-typed MCP tools for the most common operations. One MCP tool per intent, with structured inputs and validated arguments.
A non-linear DAG pipeline engine, QGIS model import, a broad vector surface, a bundled geoid, and profile-gated discovery — five ways the architecture extends beyond the operation catalog.
The GPU-resident pipeline extends from linear chains to arbitrary directed acyclic graphs. Three executors — raster, vector, and mixed — run branching workflows where a stage fans out to multiple consumers and multiple inputs converge on one operation, with raster intermediates staying resident on the device across the graph. A SavedPipeline v2 wire format carries typed slot references, conditionals, and nested pipelines.
A dedicated module reads QGIS .model3 processing models and compiles them to
ForgeGIS DAGs, with a compatibility-report mode that names any unsupported algorithm up front.
Existing QGIS investments run on the GPU-resident engine without a rebuild — the expression
layer is a clean-room QGIS Processing expression interpreter.
A broad vector surface: boolean overlay (difference, intersection), dissolve, Voronoi polygons, variable buffering, field calculators, and a clean-room expression interpreter — closing the full raster↔vector handoff grid.
The EGM96 geoid ships with the library, so orthometric/ellipsoidal vertical-datum conversion works with nothing else to download — relevant wherever elevation reference frames matter to the product.
The MCP server advertises a compact core surface by default, keeping an agent's tool list
inside context and rate-limit budgets. Opt into the full typed surface with
--profile standard; hidden tools remain callable by name — the profile governs
discovery, not capability.
GPU simulators exist; what a running ForgeGIS twin adds is the ability to branch one mid-event, and a recording that proves what you were shown. Demonstrated over real terrain around Hoover Dam: a routed release paused mid-event and forked at ten times the rate — and the gap that opened between the two worlds is exactly what the closed-form law predicts.
State advances in deterministic epochs on the GPU. A run can be paused mid-event and inspected rather than restarted from the beginning.
The branch anchors to its parent instead of copying it, so it inherits every epoch of history. Branches differ by parameters only — everything they disagree about afterwards is attributable to the value you overrode.
One conserved ledger across the whole corridor means the gap between two worlds is a number you can test against a closed-form prediction, not a picture you have to trust.
Benchmarked against GDAL CLI on equivalent workloads. Numbers are summary — full methodology, dataset descriptions, and per-operation timings live in the Technical Brief.
Full methodology, hardware configuration, and per-operation breakdowns: ForgeGIS Technical Brief (PDF).
The figures above are the GPU raster surface. The vector and topology surface is pure Java and runs on the CPU — so we measured it the same way, on real data, and published where it stops.
Microsoft building footprints for Houston through a chain of real operations — reproject, shape metrics, setback buffer, points-in-polygons join, classify — at about 125,000 features a second.
Throughput stays within about a factor of two between a hundred thousand and a million features, because every intermediate stays in process as an immutable handle and there is no per-stage file I/O term to grow with N.
The next tier up is refused before a byte is read, with the memory arithmetic in the message. A scale curve that hides where it stops is a marketing curve, so this one states it.
A long run of operations on real terrain — Copernicus elevation over the Grand Canyon and the Alps, and a Sentinel-2 scene — side by side with GDAL. Wall-clock on both sides, the running speedup along the bottom, no cuts.
gdal_calc.py invocations include Python interpreter startup × N invocations, as a real GDAL user would pay.NVIDIA RTX 5070 Laptop GPU · Intel Core 7 240H · Windows 11 · Java 17 · ForgeGIS 1.1.0 · GDAL 3.12.1. Results specific to this stack.
Per-operation JSONL timings from this recording, and the reproducible benchmark harness, available to vetted buyers under NDA — rich@seaglassfoundry.com.
Cop30 DEM contains modified Copernicus DEM data © DLR e.V. 2010–2014, Airbus DS GmbH 2014–2018, provided under COPERNICUS by the European Union and ESA. Sentinel-2 L2A contains modified Copernicus Sentinel data, ESA. GDAL is open-source software developed by the GDAL/OGR Project under an MIT-style licence.
Three audiences, three one-pagers. Each leads with the parts of ForgeGIS that matter most to that reader.
No first-party native code, no GDAL native bindings, no separate native install step — deployable in regulated environments without coordinating separate native library distribution for ForgeGIS itself. Commercially licensed; evaluation builds available.
Download defense one-pager (PDF)Drop ForgeGIS into an LLM agent platform or workflow runner and get an immediate geospatial surface: pipeline DSL for composable multi-stage work, typed intent tools for single calls. Nearly every catalog operation is reachable from an agent.
Download agent one-pager (PDF)Mid-market geospatial SaaS, climate, and agtech platforms running large raster and vector catalogs on CPU-bound GDAL pipelines can replace the compute path without migrating off the JVM or adding a native build pipeline.
Download general sales one-pager (PDF)What the engine covers. The full catalog, operation by operation, is in the Technical Brief.
Relate operation, and the raw
intersection matrix per geometry-type pair. Few GPU libraries surface the full DE-9IM
relate matrix — most stop at named predicates.
All ForgeGIS collateral. Direct download — no email gate, no form wall.
Complete operation catalog, validation methodology, benchmark details, architectural deep-dive. The substantive document for technical evaluators and engineering reviewers.
Download PDFFront/back overview. General-purpose first-touch document covering positioning, headline numbers, and the dual-surface architecture at a glance.
Download PDFFor defense and intelligence contractors, prime integrators, and SI shops. Leads on the regulated-deployment story, commercial licensing, and evaluation access.
Download PDFFor AI agent builders, LLM agent platforms, and workflow/orchestration runners. Leads on the MCP surface and the pipeline-DSL / typed-intent dual architecture.
Download PDFLicensing questions, evaluation access, partnership inquiries — we read every email.
rich@seaglassfoundry.com