Define the territory
Set the Tbilisi urban extent, ten named street corridors and a 240-cell planning grid.
LGM Research · Tbilisi · 24 August 2026
What should Tbilisi improve next?
LGM turned a broad city question into an evidence-backed shortlist of streets, districts and investment opportunities—then kept the data, method and limitations visible.
- 109,316
- rows and raster samples processed
- 88,639
- unique vector objects
- 10
- street corridors compared
- 18m 08s
- artifact-to-artifact workflow
Citywide evidence
From more than 100,000 records to places worth investigating.
LGM combined transport, places, land use, infrastructure, water, elevation and soil through one recursive MCP workflow.
The city is not one condition
A 1,050-metre elevation range changes every accessibility question.
Tbilisi ranges from approximately 360 to 1,410 metres within the research extent. A street standard that works in one district may fail in another because slope changes walking access, drainage, public transport and emergency response.
The research therefore did not search for one universal answer. It compared corridors and neighbourhoods, looking for repeated signals that justify a closer field or municipal study.
How the research ran
One question became a recursive research graph.
The MCP server supplied raster, vector and search tools. Codex orchestrated the calls, recovered from result limits and assembled the evidence.
Discover the evidence
Select Overture, OpenStreetMap, elevation and soil tools through one MCP connection.
Collect recursively
Split oversized queries into quadrants until every response falls below the 2,000-object limit.
Normalize the result
Reconstruct named corridors, merge fragments and remove overlapping IDs before comparison.
Compare places
Measure mapped access, activity, infrastructure, terrain and land-use signals by corridor and cell.
Return a shortlist
Publish candidate areas with evidence, assumptions and the checks required before a decision.
Evidence base
Large enough to see the city.
Structured enough to compare it.
The clean analytical pass involved at least 117 actual MCP calls. Exact historical tokens were unavailable because usage metadata had not yet been enabled.
Businesses, services and amenities
Crossings, stops, benches, kerbs and signals
Residential, commercial, industrial and redevelopment
Unique corridor segments after deduplication
Elevation samples across city and corridors
Comparable cells of roughly 1.25 × 1.67 km
Sources: Overture Maps release 2026-07-22.0, OpenStreetMap Nominatim, LGM Elevation and LGM Soil. The measured 18m 08s covers the artifact timeline—including MCP waits, local calculations, documentation and corridor checks—not pure server compute. A retrospective clean-pass estimate under the later lgm-0.9 rate is approximately $12–14.5.
Most important district signal
Nadzaladevi–Temka is not a single-use opportunity.
The same territory repeatedly surfaced across retail access, district employment, education and sport. That makes a coordinated mixed-use intervention more compelling than four disconnected projects.
- 01Daily retailFresh market, pharmacy and household services close to homes and stops.
- 02District jobsCoworking, banking, medical, professional and municipal services.
- 03Social infrastructureVerify school, kindergarten and sports capacity before selecting a facility.
Where to investigate next
Four decisions.
Four different geographies.
A useful city model should not produce one generic heatmap. It should surface a different evidence set for each decision.
01Daily life
Nadzaladevi–Temka
A repeated signal across daily retail, district services, education and sport makes this the strongest multi-factor priority.
02Education & sport
Gldani–Mukhiani
The strongest combined gap signal for schools, kindergartens and multifunctional sports infrastructure.
03Jobs & logistics
Varketili–Lilo
A natural fit for freight, light industry, last-mile operations and the B2B offices that support them.
04Renovation screening
Didube–Nadzaladevi
Housing near brownfield and construction activity warrants a technical audit before address-level intervention.
Street screening
Five corridors that deserve closer inspection.
These are comparative mapping signals, not proof that physical infrastructure is absent. Every recommendation begins with a field audit.
| Rank | Corridor | Sidewalk / motor | Crossings | Cycleway | Places | First move |
|---|---|---|---|---|---|---|
| 1 | Tsereteli Avenue | 6.5% | 2.38/km | 0.20 km | 518/km² | Fast field-audit pilot |
| 2 | Kakheti Highway | 3.2% | 0.39/km | 0 km | 56/km² | Reduce the barrier effect |
| 3 | Gorgasali Street | 18.2% | 0.47/km | 0.81 km | 70/km² | Reconnect east and centre |
| 4 | Ketevan Tsamebuli Ave. | 22.7% | 2.57/km | 0.21 km | 281/km² | Close walking and cycling gaps |
| 5 | Chavchavadze Avenue | 26.7% | 4.35/km | 1.67 km | 873/km² | Step-free access on steep terrain |
Tsereteli Avenue
Audit sidewalk continuity, accessible crossings and links to stops, schools and markets. Its relatively flat corridor makes rapid universal-access improvements easier to test.
Kakheti Highway
Test safe crossing intervals, continuous shaded walking routes and access to public transport. The primary question is how to reduce a long car-oriented barrier.
Example decision · Public market
Three small zones for a 10–20 stall market study.
These bounding boxes are deliberately small enough for land, footfall, ownership and field checks. They are candidate search areas—not selected plots.
0144.815–44.830 E · 41.700–41.715 N
Chugureti–Avlabari
51 residential polygons · 27 stops · 1 mapped retail-supply object
Distributed street retail: fresh food, pharmacy, household services and cafés.0244.800–44.815 E · 41.730–41.745 N
Nadzaladevi–Temka west
102 residential polygons · 31 stops · 13 retail-supply objects
A compact district centre with a 10–20 stall fresh market and daily services.0344.695–44.710 E · 41.730–41.745 N
Didi Dighomi north-west
46 residential polygons · 9 stops · no mapped retail supply
Daily-needs cluster: grocery, pharmacy, delivery, childcare and household services.From evidence to interface
The result remains inspectable and reusable.
A user can start with a city question, inspect the selected tools and sources, then continue with map-ready geometries in GIS instead of treating the answer as opaque text.
- Sources, assumptions and limitations stay attached
- Candidate areas can continue into GIS and field validation
- Future runs can attribute usage to a request and API key
What the test proved
MCP can already support autonomous city screening.
Tool discovery, recursive subdivision, vector and raster collection, street reconstruction and evidence assembly all worked in one agent workflow.
See the MCP Server →What improves next
- Server-side clipping, spatial joins and nearest-neighbour operations
- Named corridor resolution, network catchments and slope along route
- Request manifests with duration, rows, compute, tokens and cost
- Dataset version, freshness, provenance and completeness metadata
Before capital is committed
The shortlist must meet the city.
Mapped absence is not physical absence. The next stage must add population and footfall, traffic and crashes, transit frequency, property and land data, building condition, school capacity, climate exposure and on-site validation.
Soil-depth responses also require validation: the 5, 30 and 100 cm requests returned identical files. That transparency is part of the product—LGM should show not only what the evidence suggests, but also what it cannot yet support.
Bring your territory