SHEARLINE

SHEARLINE

The severe-weather analyst your agent doesn't have. SHEARLINE is a free, MIT-licensed MCP server that gives AI agents analyst-grade US severe-weather tools: live warning polygons with Impact-Based Warning tags, SPC convective outlooks, RAP-derived point environments and forecast trends (CAPE/shear/SRH/STP computed with MetPy), MRMS radar-derived hail and rotation products, GOES lightning, ground-truth storm reports (real-time and historical), and a composite threat brief that synthesizes all of it. A dozen weather MCPs already wrap the basic forecast API; SHEARLINE deliberately skips everything they do and ships only what requires radar meteorology to expose correctly.

Informational only. Not a substitute for official NWS warnings. Every tool repeats this, because it matters: when weather threatens, follow official warnings from weather.gov and local authorities.

Tools

Every tool returns structured JSON with data (numeric fields, units stated), interpretation (plain-language analyst sentences), degraded (which upstream sources failed, if any — partial data instead of errors), the safety disclaimer, and a schema_version — the field contract is semver-stable (see ARCHITECTURE), so integrators can depend on it.

Example: threat brief during a real outbreak

Real output from 2026-06-10, point inside an active tornado warning in northern Missouri:

{

"threat_level": "extreme", "threat_logic": [ "Tornado Warning in effect at the point, corroborated by confirmed tornado reports nearby — treat as an immediate life-safety situation.", "Severe Thunderstorm Warning at the point tagged 'Considerable' (hail to 1.75\", gusts to 60 mph).", "Significant-tornado parameter of 4.0 with storms ongoing — environment strongly supports tornadic supercells.", "MRMS MESH of 2.3\" hail within radius in the last hour.", "Intense rotation track (azimuthal shear 0.013 /s) nearby in the last hour.", "6 tornado report(s) near the point in the report window." ], "hazards_ranked": [ {"hazard": "tornado", "level": "extreme"}, {"hazard": "hail", "level": "extreme"}, {"hazard": "damaging_wind", "level": "extreme"}, {"hazard": "flash_flood", "level": "moderate"} ], "nearest_storm_signature": { "signature": "composite reflectivity", "value": "58.5 dBZ", "distance_km": 18.0, "direction": "ENE", "valid_utc": "2026-06-10T22:14Z" }, "attention_window": {"window": "now", "until_utc": "2026-06-10T21:00:00-05:00"} }

And the same tool for a quiet coastal Maine point reads as confidently quiet — not as an error: "threat_level": "none" with the environment numbers shown so the agent can see *why* it's quiet.

Install

Requires Python 3.12+ and uv. No API keys — every data source is public and anonymous. uvx downloads and runs the published package in one step; nothing is installed permanently.

Claude Code:

claude mcp add shearline -- uvx shearline

Claude Desktop (claude_desktop_config.json):

{

"mcpServers": { "shearline": { "command": "uvx", "args": ["shearline"] } } }

Streamable HTTP (for remote/agent-platform use):

uvx shearline --http --port 8741

serves at http://127.0.0.1:8741/mcp

On macOS you can also double-click scripts/launch.command to start the HTTP server in a Terminal window (honours SHEARLINE_PORT).

The HTTP transport is built to be hosted: it emits one structured JSON log line per tool call (tool, coarse 1° lat/lon bucket, latency, degraded list, cache hit/miss) and applies a per-client token-bucket rate limit, returning 429 with Retry-After when exceeded. Both are HTTP-only — stdio behaviour is unchanged. Configure via environment variables:

To run the latest unreleased main instead of the PyPI release, swap shearline for --from git+https://github.com/lostnumber07/shearline shearline.

Why these tools

A forecast API tells you it might rain. None of the questions that matter on a severe weather day — *is this storm rotating, how big is the hail, is the environment loaded for tornadoes, am I inside the polygon* — are answerable from a forecast endpoint. They require the warning's IBW tags, radar-derived products, and a real sounding:

• Warnings with IBW tags, not just warning text. A base-tier Severe Thunderstorm Warning and one tagged DESTRUCTIVE with 80 mph gusts are different planning problems. SHEARLINE parses the machine-readable tags (max hail size, max gust, tornado detection/damage threat) and the storm-motion vector, and does the point-in-polygon test for you.
• The environment, computed honestly. CAPE without shear is a pulse-storm day; shear without CAPE is wind-driven rain. SHEARLINE pulls the current RAP analysis profile and computes the discriminating quantities with MetPy — including the effective inflow layer, effective SRH/shear, SCP, and STP — because high-CAPE/low-shear, low-CAPE/high-shear, and classic supercell parameter spaces produce very different hazards, and the interpretation says which one you're in.
• MRMS, because warnings lag storms. MESH tells you what hail a storm has *already* produced; rotation tracks show where mesocyclones have tracked in the last hour — both on a ~2-minute cadence from the national radar mosaic, often ahead of the next warning update.
• LSRs, because radar isn't ground truth. Spotter reports confirm what's actually reaching the ground — in real time, or for any past date (the insurance / forensic question, "what hit this address that day").
• Lightning, because it's the most common outdoor-safety trigger. GOES GLM total lightning, distance-tiered to the 30-30 / 10-mile rules — the signal that actually stops a ballgame, a job site, or a drone flight.
• **Now *and* next.** The point environment is the analysis now; the forecast *trend* (f00→f06 from one model cycle) says whether STP/CAPE is rising into the afternoon — the difference between "what is it" and "is it getting worse."
• One brief that reasons across all of it. The threat level is rule-based with the triggered rules quoted back, so an agent can audit the logic instead of trusting a vibe.

Data sources (all public, no keys)

• Warnings: api.weather.gov (NWS)
• Outlooks: Storm Prediction Center public GeoJSON
• Point environment: NOMADS RAP grib filter, derived with MetPy
• MRMS: NOAA MRMS on AWS Open Data
• Storm reports (real-time and historical): Iowa Environmental Mesonet LSR service
• NEXRAD Level 2: Unidata on AWS Open Data
• Lightning: GOES GLM on AWS Open Data (GOES-East GLM-L2-LCFA)

Coverage is continental US only — out-of-bounds coordinates are rejected with a clear error. Upstream fetches are cached (warnings 60 s, MRMS/lightning 120 s, LSRs 300 s, outlooks/RAP 30 min, historical reports 6 h) and degrade gracefully: if one source is down, you get partial data plus a degraded field, never a bare exception. A daily canary workflow checks every upstream's response shape and fails on schema drift (renamed fields, moved buckets) before it reaches you.

Recipes for non-meteorologists

You don't need to know what an STP is to use SHEARLINE. The .claude/skills/ directory ships three end-to-end recipes that name the exact tool sequence for a domain task — drop them into any agent that has SHEARLINE connected:

• [hail-claim-verification](.claude/skills/hail-claim-verification/SKILL.md) — did damaging hail occur at this address on this date? (insurance / forensic)
• [chase-day-briefing](.claude/skills/chase-day-briefing/SKILL.md) — outlook → environment → trend → warnings → radar, into a go/no-go with a target