New · Forge Suite · v1.2.2

ForgeGIS Studio

Where you see the Forge suite work — and shape it to your mission

ForgeGIS computes, ForgeData catalogs, and ForgeMind reasons — but each speaks only libraries and machine protocols. Studio is the operator surface that unifies them: one browser workspace where data is loaded, pipelines are composed and run, and a natural-language agent acts on a live map.

1 host
Whole suite runs on a single workstation — loopback-bound, air-gap capable
3 engines
GPU compute, spatial catalog, and a natural-language agent behind one surface
R · V · M
Raster, vector & mixed pipelines — non-linear DAGs authored and run in batch
The ForgeGIS Studio workspace: a GPU-computed hillshade rendered over the live map of southeastern Arizona, with the catalog rail of SRTM elevation tiles on the left, two labeled points on the map, and the Inspector panel showing the hillshade operation's display settings
The Studio workspace: a GPU-computed hillshade of an SRTM elevation tile rendered on the live map over southeastern Arizona, with two named points placed as scene entities. The catalog rail (left) browses ForgeData's elevation tiles; the Inspector (right) shows the hillshade operation's colormap, opacity, CRS, and extent; the status bar reports the GPU device, heap, and live handles.

What it is, why it exists, how it's built differently

Three short answers for the technical evaluator.

What it is

Studio is the browser workspace that puts a live map, a pipeline canvas, and a conversational agent in front of a person, so the suite's capabilities can be seen, run, and evaluated through one interface. It is a demonstration and operator surface — not a general-purpose GIS desktop, and it does not try to be one.

Why it exists

The engines are deliberately headless — libraries and MCP servers with no UI. That keeps each engine focused and independently deployable, but it means a person needs somewhere to stand. Studio is that place: the one component whose job is to make the suite usable and visible, so the engines never have to grow a UI each.

How it's built differently

Studio is thin over stable engines on purpose. Operations flow through from the ForgeGIS catalog and render as nodes automatically; workflows are saved artifacts, not code; the agent's on-map actions extend through a clean bridge. Tailoring the surface to your program is composing known pieces — not building a platform.

Compose pipelines by hand

A node editor where ForgeGIS operations become nodes. Raster, vector, and mixed raster-to-vector graphs — with branches and joins — are authored on the canvas, run, saved to a library, and replayed. Topology, not a single linear chain.

ForgeGIS Studio pipeline canvas: a DEM source node connected to a GPU hillshade operation node, each with a live result thumbnail
The pipeline canvas: a DEM source node feeding a GPU hillshade operation, each with a live result thumbnail. Graphs built here save to a library and replay — on the canvas or in batch.

Type-valid by construction

The palette gates operations by the selected node's kind — vector operations off a vector node, a raster-to-vector operation off a raster node — so the graph the operator builds is always type-valid. A Polygonize → Buffer chain that crosses from raster into vector is expressed directly.

A format that carries topology

Pipelines serialize to a versioned format that records the pipeline type, each operation's kind, its inputs, and its dependencies. One pipeline representation serves interactive and batch execution — not two.

Run one, then run hundreds

The batch runner executes the same saved pipelines across a folder of inputs — so a graph proven on one file scales to production without a memory surprise.

ForgeGIS Studio batch tab: configure a saved pipeline over filesystem or catalog inputs, with output directory, error handling, and a live per-tile log
Configuring a batch: a saved pipeline, filesystem or catalog inputs, an output directory, and error handling — per-tile completions stream into the live log once it runs.

Resource pre-flight, not trial and error

Before committing a batch, Studio estimates each tile's peak VRAM through the shared resource-estimation contract exposed by ForgeGIS, and gates the run against the available budget rather than discovering the ceiling by crashing. Interactive applies use the same estimator — one memory ceiling across both.

Built to run unattended

A priority queue, progress events per operation, out-of-bounds inputs skipped rather than failing the whole batch, and a CSV report with estimated and actual VRAM per tile so a run can be audited after the fact.

Let the agent drive the map

The most distinctive thing Studio does: a natural-language agent acts on the operator's own map — not a copy.

Two paths, one deliberately more capable

Over the plain chat rail, ForgeMind answers in text and runs read-only operations. To act on the canvas — place markers, draw overlays, push reports — the agent connects through the studio-mcp bridge, an authenticated channel that exposes Studio's map as MCP tools. Reading the map is lightweight and always available; driving the map requires the bridge and its credential.

Work that persists

Named artifacts the agent places on the map are recorded as durable scene entities in the ForgeData catalog — the source of truth, not session scratch. Restart the suite and the agent's cached handles still resolve. Work is not lost when the process cycles.

The ForgeMind chat rail inside ForgeGIS Studio, showing a conversation that reads and acts on the live map
The ForgeMind chat rail: a question about the current view, the studio.get_viewport and forgedata.elevation_in_region tool calls it triggered, and an answer grounded in real numbers.

Data through one catalog

Studio reads through ForgeData rather than owning file I/O — one catalog and one extraction strategy across the suite.

Catalog-backed loads

An operator browses ForgeData's registered datasets and loads one into the session. Studio does not reimplement the catalog's knowledge of where data lives or how to reach remote tiers — it routes through it.

AOI-windowed reads

Supply an area of interest and Studio asks ForgeData for just that window rather than the full source tile — a canyon's worth of pixels, not a continental raster. The windowed extract is sized to fit under Studio's raster cap before it ever reaches the GPU.

Presence verification

A catalog can register more than is actually on disk. Studio consumes ForgeData's presence-verification signal so a registered-but-missing dataset is caught at pick time with a clear message — not at read time with an opaque file-not-found.

Security posture

Single-user, single-host is a security decision, not a limitation — and it is enforced, not assumed.

Loopback by default

The server binds to 127.0.0.1

Other hosts on the network cannot reach it. Rebind outward and Studio warns on every boot until hosted-deployment prerequisites — auth, CSRF, rate limiting, TLS, audit — are in place. The warning is deliberate.

Mediated paths

A sandbox on every path route

Every route accepting a caller-controlled path runs it through a sandbox. Uploads are confined and extension-checked; user-added roots are canonicalized and rejected if they reach sensitive locations; agent tool calls run operator-only.

Bounded resources

The worst case is capped

Upload, pixel, and dimension caps plus a pre-dispatch VRAM gate, a bounded op-executor pool that refuses rather than thrashes, and per-operation timeouts bound the worst case a single workstation must absorb.

A separate door

The agent has its own credential

The browser session is safe only under loopback. The agent reaches the map through a distinct, token-validated bridge with an agent-key exchange — the stronger credential guards the ability to act on the map.

On an unclassified on-premises network, the whole suite runs on one host, reachable only from that host's own browser, with no external calls required for the core workflow — including on a disconnected network.

Built to be customized, fast

The compute, the catalog, and the natural-language control plane are the hard, stable parts of the suite. Studio is the layer that decides what a particular customer sees and touches — and re-cutting that layer does not disturb anything underneath.

Capabilities flow through

Operations are not enumerated by hand in Studio; they arrive from the ForgeGIS catalog, stamped with kind and input ports, and render as nodes automatically. When the engine gains an operation, it appears in Studio without a Studio code change.

Workflows are data, not code

A workflow tuned for a customer is a saved artifact that ships with a deployment and replays faithfully, on the canvas or in the batch runner. Curating a library of mission-specific pipelines is a configuration task, not an engineering one.

The agent surface composes too

The studio-mcp bridge exposes Studio's map to ForgeMind as MCP tools. Extending what the agent can do on the canvas — a new overlay, a customer-specific report — is adding a tool to that bridge, cleanly separated from both engine and browser UI.

A customer engagement typically does not begin with “build a GIS application.” It begins with a working suite and a working Studio, and the work is shaping the surface — views, palette, curated workflows, agent actions — to the customer's task. Seaglass Foundry turns that customization around directly.

Where Studio sits in the suite

Studio is the operator surface of the Forge suite and is not sold separately. Its value is realized with the engines behind it.

ForgeGIS
Computes
GPU-accelerated geospatial compute engine — raster, terrain, spectral, point clouds.
ForgeMind
Reasons
The GIS Intent Compiler — natural-language orchestration across tools.
ForgeData
Decides the data
Catalog, cost routing, sync, derived-product memory, window extraction.
ForgeGIS Studio
The operator surface
One workspace — live map, pipeline canvas, batch runner, and agent chat over the same session.

One session, many surfaces: the browser, the batch runner, and the agent bridge all act within the same session, over the same handles. The agent draws on the map the operator is looking at — not a copy. The suite behaves as one system because Studio makes it one workspace.

Single-host today, built to scale outward

The single-host model is the first configuration, not the only intended one. Three configurations describe the vision — a stated direction, not shipping features, with the hard work named honestly.

One GPU, multiple clients

A single accelerator-equipped host serving several operators. The hard problems: fair scheduling of a shared VRAM budget across concurrent sessions, isolation between operators, and real multi-user authentication.

One client, multiple GPUs

A single operator's heavy pipeline spread across several accelerators — for the large-area, high-resolution work where one GPU's memory is the limit. The hard problems: partitioning a pipeline across devices and reconciling resource estimation with a multi-device budget.

Multiple clients, multiple GPUs

The general case: a pool of operators drawing on a pool of accelerators. Composes the first two and adds placement and scheduling across a device pool, plus the full hosted-deployment security posture Studio's design already names as prerequisites.

What makes the progression credible is the architecture Studio already has: a thin surface over engines that own their own resource accounting, and a security model explicit about exactly what must be added before the port leaves the loopback. The foundation was built to be grown from.

Downloads

All ForgeGIS Studio collateral. Direct download — no email gate, no form wall.

A note on licensing. ForgeGIS Studio is the operator surface of the Forge suite and is not sold separately. Its license is not yet finalized; nothing here should be read as a grant of rights. Contact the maintainer about evaluation access.

Evaluate the Forge suite — see it work on one machine

Run the whole suite on one evaluation host, disconnected if required, and exercise every part of it — compute, catalog, and agent — through one interface.

rich@seaglassfoundry.com