Skip to main content
Verdant’s agent skills turn a research question into a reproducible analysis using our public API or data MCP. They need no handoff archive, database credentials or private source checkout. Fresh API responses can be cached for the run.

​
Install

Install the four published skills from the Verdant repository:
Choose your agent and installation scope in the CLI. You can also install a single skill by passing only its name after --skill. Installation uses the skills CLI; its default install telemetry feeds skills.sh discovery. A directory listing may take time to appear. Browse Verdant on skills.sh.

​
Choose a workflow

Each skill can be installed independently. The discovery skill includes an optional Python standard-library downloader that verifies full dataset content hashes and records HTTP provenance. The domain skills guide independent calculations; they do not ship scientific datasets or saved expected answers.

​
Example prompts

  • “Use $verdant-data-discovery to find evidence for grape irrigation decisions. Explain which datasets support a historical economic comparison.”
  • “Use $verdant-forecast-backtest to replay day-two rain and frost protection from the service. Report every declared threshold.”
  • “Use $verdant-perennial-economics to compare all CSIRO treatment years at crop price 0.50/kg and water value 1.00/m³, equal quality and zero extra cost.”
  • “Use $verdant-nitrogen-replay to rescore every Ohio variant and show both equal-year and equal-trial margins.”

​
Connect and reproduce

Connect https://api.verdant-ai.com/mcp in a Streamable HTTP client, or use the public REST API directly. See MCP connection instructions. Installing a skill does not itself register an MCP server. Current public reads require no credentials. Skills discover the live catalog, inspect methodology, pin immutable versions, validate complete cohorts and compute from inputs. They preserve unfavorable seasons and variants, and distinguish measured outcomes from conditional financial assumptions. Save the calculation code, request receipt, assumptions, per-year results and frozen output digest with each run. The service supports frozen forecast/policy replay. Fresh NWS calibration and full Ohio refitting need additional inputs or specifications. Current acquisition is disabled: a missing-data brief is not a queued data request. Catalog reference scores can expose expected results during discovery; disclose this before calling an evaluation blinded. These historical analyses do not establish prospective farmer-profit gains.