The full upstream README, mirrored here for reference. Install config, tool schemas, adoption signals, and an original overview live on the Solar Data listing page.
US solar data, agent-accessible.
One MCP server that brings US open solar data to Claude, ChatGPT, and anything else that speaks MCP.
Install one server — uvx solar-data-mcp — and your agent gets all 18 tools, 11
skills that teach it how to use them, and 4 ready-made report prompts, across four
data domains: production modeling, economics, market data, and forecasts. Every tool
returns the same envelope — data + units + source + assumptions + warnings —
so the agent always knows what a number means, where it came from, and which defaults
were injected on its behalf.
Behind each question the agent picks the right tools (or a skill routes it), and every number comes back with units, a source, and the assumptions made on your behalf.
Get a free NREL API key — https://developer.nlr.gov/signup/
(or use DEMO_KEY to try it out — 10 requests/hour).
Add the server to your agent — snippets for every major agent below; for
Claude Desktop, merge this into claude_desktop_config.json
(full example in examples/):
Only NREL_API_KEY is needed to start: the server runs with any subset of keys,
and a tool missing its key returns setup instructions instead of failing silently.
Restart your client and ask — try "Compare annual production for an 8 kW system in Mesa, AZ at 10° vs 25° tilt."
Verify keys and connectivity anytime:
Claude Code
(or commit the quickstart JSON to your project's .mcp.json)
Codex CLI — ~/.codex/config.toml:
(or codex mcp add solar-data --env NREL_API_KEY=YOUR_KEY -- uvx solar-data-mcp;
use env_vars = ["NREL_API_KEY"] to forward keys from your shell instead of
hardcoding them)
OpenCode — opencode.json in your project (or ~/.config/opencode/opencode.json):
Hermes — ~/.hermes/config.yaml, then /reload-mcp:
Claude Desktop — the quickstart JSON above.
Anything else that speaks MCP (stdio) — command uvx, args ["solar-data-mcp"],
keys in the env block.
All free, and every one optional — the server starts with none set.
| Env var | Unlocks | Get one |
|---|---|---|
NREL_API_KEY | estimate_production, get_solar_resource, compare_orientations, size_system_for_target, estimate_roi, compare_forecast_to_model | https://developer.nlr.gov/signup/ |
OPENEI_API_KEY | lookup_tariffs | https://openei.org/services/api/signup/ |
EIA_API_KEY | get_electricity_prices | https://www.eia.gov/opendata/register.php |
AHJ_REGISTRY_TOKEN (optional) | identify_ahj | email support@sunspec.org |
Market tools (USPVDB, Tracking the Sun, SolarTRACE) and forecasts need no key.
Full forecast output additionally needs the Quartz model installed into a persistent
environment (see packages/solar-forecast/);
without it the forecast tools return install instructions.
uvx solar-data-mcp — the install above — serves all four domains on one stdio entry.
Each domain also ships as a standalone server:
| Domain | Data | Standalone server |
|---|---|---|
| Production | PVWatts v8 modeling, NSRDB irradiance | uvx --from solar-data-mcp-nrel nrel-solar-mcp |
| Economics | URDB tariffs, EIA prices, federal ITC + DSIRE | uvx --from solar-data-mcp-economics solar-economics-mcp |
| Market | SolarTRACE, Tracking the Sun, USPVDB, AHJ lookup | uvx --from solar-data-mcp-market solar-market-mcp |
| Forecast | Quartz open-source forecasts (OCF) | uvx --from solar-data-mcp-forecast solar-forecast-mcp |
Per-server config:
examples/claude_desktop_config.per-server.json.
⚠️ Run the combined
solar-dataserver or the per-domain servers, not both — and neversolar-economicsandsolar-marketside by side. Both open the same local DuckDB bulk store, which allows only one process at a time.
Eighteen tools across four domains; parameter details live in each tool's docstring and each domain package's README.
| Domain | Ask it about | Tools |
|---|---|---|
| Production | output, sunniness, sizing, roof orientation | estimate_production, get_solar_resource, compare_orientations, size_system_for_target |
| Economics | tariffs, electricity prices, incentives, payback | lookup_tariffs, get_electricity_prices, get_incentives, estimate_roi, sync_incentives |
| Market | installed $/W, permitting times, utility-scale plants | query_installed_systems, get_permitting_timelines, find_utility_scale_projects, identify_ahj, market_snapshot, sync_* |
| Forecast | tomorrow's output, "is today unusual?" | forecast_generation, compare_forecast_to_model |
Skills are procedures shipped inside the combined server that teach an agent to chain
the tools correctly — ordering, sync prerequisites, honest reporting. They're MCP
resources: skill://solar/index routes by question shape, skill://solar/<name> is
the procedure. Grouped by who's asking:
Four of these render reports with a fixed document shape and are also exposed as
MCP prompts your host surfaces natively — market_brief, site_assessment,
quote_review, proposal_builder (in Claude Code: /mcp__solar-data__market_brief).
Full catalog, routing design, and report templates: docs/skills.md.
Layout: packages/core (shared HTTP client, cache, envelope), one package per domain
server, and packages/solar-data-mcp (the umbrella that mounts all four on one stdio
entry). Smallest possible client:
examples/example_client.py.
MIT. Per-source data licensing/attribution is exposed as MCP resources
(source://<name>/license).