Windows desktop · LiDAR + AI · early alpha
Your survey.
Your AI assistant.
Your control.
View, edit and review large LiDAR point clouds. Connect Codex, Claude or another compatible AI app through MCP to inspect your survey and make undoable edits. Choose what it can change, then check the results in the same desktop workspace you use yourself.
Designed and developed in the United States →
Bring your AI assistant into Pointglass
Let AI see the cloud.
Keep the decisions yours.
Ask a compatible assistant such as Codex or Claude to inspect the survey you have open, investigate a region and help with the edits you authorize.
Pointglass uses the Model Context Protocol (MCP) to give your connected AI app views of the cloud and tools to work with it. The assistant can move around the survey, compare colour modes, probe points and check regional statistics before proposing a change.
- Inspect with context. Captures, a reference grid, colour sampling and bounded queries help the assistant investigate specific areas.
- Work on a defined scope. Select by region or attributes, isolate verified Terra flights, and preview supported classification or trim plans.
- Review what happened. AI edits appear in ordinary edit history with attribution. Undo an edit, replay the timelapse or revert an AI session.
“Inspect this area for possible ground-classification omissions. Show me the evidence and preview the points before changing anything.”
An example instruction; AI interpretations and proposed classifications need review.Disconnected
The default at startup. AI cannot connect until you enable control, unless you explicitly choose a launch option.
Inspect together
Let the assistant view, navigate, filter and select. It cannot change points.
Approve each edit
Review the proposed operation, its point count and scope in the AI panel before it runs.
Set the boundaries
Choose point limits, permitted actions, stores and an expiry. Work outside those limits still asks for approval.
Connect once. Choose the level each session.
Open Settings → Connect AI apps… to preview setup changes. Enable AI control separately, then use the AI panel to follow activity, stop work or restore your view. AI editing uses the same editing licence as manual edits.
While AI control is enabled, survey views and details are shared with your connected AI app and may be sent to its provider. Manual Pointglass workflows work without an AI provider. Supported MCP tools cover specific workflows; the bridge does not automate every desktop feature.
One workspace for the whole survey review
Find the flight behind the points
Link multiple Terra projects, inspect trajectories and flight metrics, and isolate verified flight points in any colour mode. Flight visibility and undoable point deletion are separate actions.
Build rasters from the cloud
Create ground elevation, top-surface elevation and RGB rasters for a store, view or measurement area. Export GeoTIFF with progress and cancellation. RGB comes from point colours.
Keep findings on the map
Place coloured pins with titles and descriptions, revisit them from a list and save them with the project. Include an RGB overview, coordinates and close-ups in reports.
Compare the model with LiDAR
Attach meshes with explicit CRS, units and offsets. Compare by overlay, blink, opacity or wireframe. Imported meshes remain read-only; point editing applies to LiDAR.
Review roofs, export outlines
Preview roof candidates before applying undoable Building-class edits. Export DXF outlines and an optional LandXML ground TIN, with results reviewed before delivery.
Narrow down the evidence
Sample colours to include, exclude or select matching points. Use bounded terrain diagnostics to investigate points above supported ground, with uncertainty shown separately.
Terra QA also provides a standalone workspace for acquisition, terrain, overlap and survey checks. Terrain diagnostics do not automatically identify rocks, reclassify points or produce a recovered surface.
A cross-section is the fastest way to answer a question about a cloud
Draw a corridor in the map and inspect it side-on in its own window. The extractor descends the octree independently of the viewport render budget, extracting the corridor at full density and reporting when its point cap is reached.
Walk the corridor along its axis, or rotate it 20° at a time about its midpoint to square the section up to a feature. The vertical exaggeration is a control, not a guess, and the elevation line reads in the project's own units.
- Full density. 182,127 points in the section on the right, from a 217-million-point store.
- Its own window. Pop it out to a second monitor and keep orbiting the map.
- Measure inside it. The section is a working view, not a preview.
Every tree tip, where it was measured
Import an obstruction CSV and each tip lands in the cloud at its own coordinates. Columns are matched by alias rather than by position, so a file that names a column tip_elev and one that names it Tip Elevation (ft) both load.
Sort by tip elevation, height above ground, tier or zone. Mark rows and they are drawn in the viewport. Set a centrepoint — a radome, a runway end, a proposed structure — and read the 3D straight-line distance to each selected tree tip. The calculation uses easting, northing and tip elevation, with labels in the project's units.
- The table pops out. Full-height on a second monitor, sorted numerically — the proxy sorts on the raw value, not the display string.
- Boundaries overlay the cloud. Parcel and zone vectors project onto the surface at an anchor elevation you set.
- Tip elevation matters. Each tree-to-centrepoint distance includes the vertical separation from the dome reference.
Ground or roof? Answer it by looking
Measurements are drawn straight onto the cloud with live dimension labels. The first click snaps its elevation to the ground-classified surface beneath it — so a circle centred on a radome sits on the ground under the dome, not on its roof — while keeping its XY exactly under the crosshair.
Picking the circle tool snaps the view to a framed top-down plan and holds it there. While the centre is live, an inset shows a vertical section of the square around the crosshair with the anchor's elevation drawn across it. Arrow keys nudge the centre across the plan; PgUp and PgDn nudge its elevation.
- Any closed shape is a trim boundary. It replays through the same pipeline a lasso does, so scope, confirmation and undo behave identically.
- Measurements are saved beside the project. A boundary worked out in one session is there in the next.
Accuracy numbers that survive being read closely
The report examples use fictional project names, people, companies, coordinates and measurements. Choose a standards-style report or plain accuracy report in the current report dialog.
Load checkpoints and GCPs, measure vertical residuals against a ground TIN, and keep independent checkpoints separate from points used to build the model. Ground support and vegetation groups remain visible in the results.
The current ground-TIN workflow reports observed RMSE against ASPRS 2024 (Edition 2, Version 2), with non-vegetated and vegetated checkpoints kept separate. It does not assign an automatic accuracy class or pass/fail verdict. A legacy local-plane method remains available with its own statistics.
Observed RMSE
Report the measured error, separate vegetation groups and show the supporting evidence without an automatic accuracy verdict.
Checkpoints ≠ GCPs
Points withheld from model construction stay separate from the control used to create it, in the table, statistics and report.
Local-plane results
Legacy reports retain NVA₉₅ and VVA₉₅. Groups below the 30-checkpoint minimum are flagged, with the method identified.
Horizontal, when you ask for it
Vertical is what a lidar survey is usually judged on, and it is what Pointglass measures by default. Horizontal is a separate switch in the Control panel, off unless you turn it on — because a horizontal error is something this software reports, never something it corrects.
You place the pick yourself. Automatic panel-finding by intensity contrast is deliberately not implemented: it locks onto a road stripe as readily as onto a target, and a confident wrong answer is worse here than no answer. Each pick is written to a .picks.csv beside the survey — never into it, so a re-export from your survey software cannot silently discard your observations.
- The 95% figure is not one formula, and it does not pretend to be. Equal axes give NSSDA's 2.4477 × RMSEx; comparable axes give the averaged form. Below the standard's own 0.6 axis-ratio floor there is no published expression — so the components and the mean shift are still reported and the 95% claim is withheld, and said to be withheld.
- The picked elevation is discarded. A pick answers where in plan. Vertical residuals come from the chosen surface method; the two measurements answer separate questions.
- Off means absent, not empty. The table and report show horizontal fields only when horizontal measurement is enabled.
- The method states its own floor. A pick cannot resolve finer than the point spacing plus the operator's aim, and the report says so in its lede.
The report is the deliverable
One command produces a client document rather than a screenshot of a table. The same numbers come out as CSV for a spreadsheet.
- Letterhead and a metadata grid — project, date, what was measured, and the coordinate system on its own row, printed whole with both EPSG codes rather than elided.
- Per-group KPI tiles — checkpoints and GCPs reported separately, always.
- A signed residual bar chart with a ±RMSE band, so sign and spread are both visible.
- The residual table, every point, no truncation.
- Horizontal accuracy and per-point offsets — only when horizontal was actually measured.
- Cloud statistics — point total, classification breakdown and returns, read from what the builder already recorded rather than by rescanning billions of points.
- Optional Qualifications and glossary sections to explain the results and their limits.
- Three full-report styles — ASPRS, Plain and DJI Terra style, with residual CSV for spreadsheets.
- Full-page image exhibits, optional marker overview and close-ups, flight sections, base-station/OPUS details and a condensed one-page option.
Editing a cloud should not change the points you did not edit
Copied, not regenerated
Every surviving point record is copied byte for byte from the original source file, with only its classification byte patched. Headers and VLRs come across intact. The one documented exception is the 48-byte header bounding box, which is recomputed — and says so.
Refuses what it cannot prove
Any declared coordinate system is read natively — the CRS and its EPSG codes are carried through from the source file to the viewport badge to the report, and the linear unit travels with them. What is refused is adding or merging a file whose CRS or unit does not provably match the project: byte-identical CRS records, or exact EPSG authority-and-unit equality. Never converted, never overridable.
Compare without merging
Add LAS or LAZ to a project after its initial build; each addition becomes its own complete sub-store, drawn alongside the others. Exactly one store is active and receives edits. Every store keeps its own history and its own export.
Thirteen panels, and none of them lie to you
Panels stack as collapsible sections in a rail down each side. A rail scrolls, so every panel keeps its natural height however many are open. A dot beside a panel's title means that panel is still filtering the cloud — a collapsed filter cannot quietly change what you see.
Any panel pops out into its own window and goes back where it was when you close it. Positions, sizes and visibility survive between sessions. Every command has a key, and every key can be changed.
- Read in feet or metres, whatever the file is in. Measurements, profiles, obstructions and the readouts all follow one display-unit setting. The store's own units never change; only what you are shown does.
- The CRS is on screen, not in a dialog. A badge under the orientation gizmo names the project's coordinate system and its unit, so the number you just read can always be attributed.
- Each store remembers its view. Camera, filters and display state are saved per store and restored when you open it again.
- The elevation ramp explains its own range. Choose the rule — percentile, median ± MAD, or the full declared box — restrict it to ground, refit it on demand, and read it against a legend drawn from the ramp itself. Noise a thousand feet below ground stops deciding what the colours mean.
A LAS file is stored in the order the sensor fired
That order is useless for interactive work, and a multi-gigabyte LAS cannot be edited in place. So the first thing Pointglass does is rewrite the cloud into a spatially indexed store that separates immutable geometry from a small mutable attribute layer. A node's points are a contiguous slice of one file, which is why drawing one is a direct upload to the GPU and why level-of-detail streaming keeps a multi-billion-point project interactive — the 4.29-billion-point store ceiling came out, and the largest survey on record is 4.82 billion.
The second decision is a boundary. Geometry, statistics, report composition and the click grammars are Qt-free modules; the widgets are reduced to wiring. A test enforces it. That is why a desktop application can carry thousands of tests, and why the command line runs on an interpreter with no Qt installed at all.
Originally designed with NDAA LiDAR workflows in mind
Designed and developed
in the United States.
Pointglass was originally designed to support the review and editing of point-cloud data collected with NDAA-compliant LiDAR payloads, using their LAS/LAZ deliverables.
Pointglass-specific product design and original application-code development are performed entirely in the United States.
This development-location statement covers Avenislabs’ own work on Pointglass. Third-party libraries and runtimes have their own origins, and connected AI services are supplied separately.
Work with the files on your workstation
Viewing, editing and report generation run in the Windows desktop application. Licensing, update checks, enabled diagnostic uploads and optional AI connections have separate network activity.
Choose AI to match your project’s rules
Manual workflows require no AI provider. When enabled, AI control shares survey views and details with the connected AI app and its provider. Review that service’s access, retention and data-location terms before using it on restricted work.
What this means for NDAA-related procurement
U.S. development and compatibility with a payload’s exported files are facts about the software’s design and use. Procurement eligibility depends on the applicable law and contract, the complete hardware and software configuration, suppliers and ownership, and network and storage arrangements.
Pointglass makes no blanket “NDAA-certified” or government-approval claim. Processing a file does not certify its source payload or aircraft. A project-specific assessment must also consider third-party dependencies and any AI service used.
Read the relevant federal requirements and origin-claim guidance
- LiDAR: FY2025 NDAA §164, as amended by FY2026 NDAA §162 — DoD restrictions address covered LiDAR suppliers, operating-software origin, and network or storage arrangements. The statutory prohibition took effect June 30, 2026.
- UAS: FY2020 NDAA §848, as amended by FY2023 NDAA §817 — separate restrictions concern aircraft systems, related equipment and services, software and covered suppliers.
- American Security Drone Act / FAR 52.240-1 and Section 889 / FAR 52.204-25 — additional UAS and telecommunications restrictions may apply to the procurement and contractor.
- FTC guidance on U.S.-origin claims — specific design and development statements must be truthful and substantiated; broader product-origin claims require their own evidence.
Sources reviewed November 4, 2026. These references explain the scope of our statements and do not establish eligibility for an individual procurement.
This is early alpha, and it says so in the title bar
This feature overview reflects the November 4, 2026 checkout, version 0.33.56. Targeted tests and selected real-data checks support individual workflows; they do not establish survey certification or full acceptance of every feature. Installer availability follows the release feed. The version scheme is deliberately fine-grained: patch bumps are the norm and a large patch counter is expected. Version 1.0 will be an explicit decision, not somewhere the counter arrives.
There is no public download yet. If your work looks like the work in these screenshots, get in touch and say what you are trying to do with it.
The form below reaches me directly.
| Operating system | Windows |
| Graphics | Any GPU and driver supporting OpenGL 3.3 core — roughly 2010 onward. The window says so plainly rather than opening blank. |
| Runtime | Python 3.12+, numpy, laspy[lazrs] and cryptography for a source installation. Pointglass.bat installs the checkout dependencies. |
| Desktop window | PySide6, about 150 MB, installed on first run. |
| Optional | Report, CRS and QA dependencies for their respective workflows; the MCP bridge for external AI control. Manual operation requires no AI provider. |
Say what you are trying to do with it
There is no public download yet, and access is handled one conversation at a time. The most useful thing you can tell me is the shape of your work: how large the surveys are, what you deliver, and which part of it is currently painful.
Nothing here is stored on the site — the message is relayed straight through and answered by email.