DEV Community 1h ago

Tell your AI agent if the ground is sinking: measured ground-motion with SibFly + LangChain

Most "is this property risky?" data your agent can reach is modeled - a prediction from soil maps and flood zones. There's a different, underused signal: measured ground motion. Satellites have been tracking how far the ground has physically moved, to the millimeter, for years.

SibFly packages that into one call: give it a US address, get back how fast the ground is sinking or rising, in mm/year, derived from NASA's OPERA Sentinel-1 InSAR dataset. Measured, not modeled. Negative = sinking. This post wires it into a LangChain agent in about 30 lines.

Why measured ground motion?

Ground subsidence (sinking) quietly wrecks foundations, and it's invisible on a normal map. It shows up in radar interferometry (InSAR): compare two satellite passes over the same spot and you can measure sub-centimeter vertical change. NASA processes this for all of North America and publishes it for free - but the raw data is 400 MB HDF5 granules in radar geometry, not something an agent can query per address. SibFly does the ETL and serves the pixel: GET /api/v1/motion?address=... → a JSON verdict.

Install

pip install -U langchain-sibfly
export SIBFLY_API_KEY="sf_live_..."

Grab a key with free starter credits at sibfly.com. Agents can also self-register with no human in the loop:

curl -X POST https://sibfly.com/api/v1/autonomous/register \
  -H "Content-Type: application/json" \
  -d '{"email":"you@example.com"}'
# -> { "api_key": "sf_live_...", "credits_usd": 1.0, ... }

The tool on its own

from langchain_sibfly import SibflyGroundMotion

motion = SibflyGroundMotion()  # reads SIBFLY_API_KEY
print(motion.invoke({"address": "1100 Congress Ave, Austin, TX"}))

{
  "status": "ok",
  "velocity_vertical_mm_yr": -6.0,
  "velocity_uncertainty_mm_yr": 1.5,
  "assessment_code": "notable_subsidence",
  "confidence": 0.86,
  "data_age_days": 73,
  "cost_usd": 0.4,
  "credits_remaining_usd": 0.6
}

Route your logic on assessment_code - a stable enum (rapid_subsidence, notable_subsidence, stable, mild_uplift, strong_uplift) - not on the human-readable string.

Pricing that doesn't punish exploration

The thing I appreciate as someone who builds agents: misses are free. Out-of-coverage, no-data, too-stale, and low-confidence responses all come back HTTP 200 with cost_usd: 0. You only pay $0.40 when you get a real covered reading.

And there's a free preview:

# free: coverage + confidence + would_cost_usd, no charge
motion.invoke({"lat": 30.3244, "lon": -97.8102, "dry_run": True})

You can also refuse to pay for data you'd reject, server-side:

motion.invoke({
    "address": "...",
    "max_age_days": 400,     # too old -> free "stale_data"
    "min_confidence": 0.7,   # too noisy -> free "low_confidence"
})

In an agent

from langchain_sibfly import SibflyGroundMotion
from langchain_anthropic import ChatAnthropic
from langgraph.prebuilt import create_react_agent

agent = create_react_agent(
    ChatAnthropic(model="claude-sonnet-5"),
    tools=[SibflyGroundMotion()],
)

out = agent.invoke({"messages": [
    ("user", "Is the ground sinking under 1100 Congress Ave, Austin TX? "
             "Check coverage first so we don't waste a call.")
]})
print(out["messages"][-1].content)

The agent will typically dry_run first (free), see it's covered, then pull the real reading.

Prefer MCP?

SibFly also runs a hosted MCP server, so any MCP client can use it with no SDK: https://sibfly.com/mcp (Streamable HTTP, Bearer auth). It's in the official MCP registry as com.sibfly/ground-motion.

Tools: check_ground_motion, check_coverage, check_portfolio, get_motion_history, get_account.

One honest caveat

This is a screening signal, not an engineering survey. It tells you a parcel is moving and roughly how fast; it doesn't replace a geotechnical report. SibFly returns an uncertainty (±mm/yr), a confidence score, and a neighbor_consistent flag so your agent can tell a real trend from a single noisy pixel. Treat it like a credit-score check: fast, cheap, directional.