An open commons of evidence, protocols, data and analysis for agricultural research
Visit the website at agrdatasci.github.io/maniva.
Kauê de Sousa, Marie-Angélique Laporte, Brayden Youngberg, Almendra Cremaschi, Rachel Chase, Jacob van Etten
Agricultural research produces papers, protocols, data and code, but they sit in separate places with nothing linking them. Using them well takes an insider, someone who knows which paper backs a method, which protocol produced a dataset and which script analyses it. Breeders in national programmes, students, modellers, other research teams and funders have to ask that insider. AI assistants fill the gap with generic or invented answers. maniva links these assets and documents them so that anyone, and any AI assistant, can use them without the insider.
maniva makes each asset AI-ready, which means each one comes with a lightweight skill that tells an AI assistant what the asset contains, how to use it and how to cite it. A user of tricot, the on-farm testing approach that is maniva's first domain, can then ask how many farms a trial needs, get the answer from our published evidence, follow the matching protocol, and load the open data behind it, all in one conversation.
Share this one link. Copy the link below, paste it into an AI assistant with web access and ask your question.
https://github.com/AgrDataSci/maniva
The assistant reads AGENTS.md and SKILL.md, which route the question to the right layer and load only what it needs. The same link keeps working as layers are added.
In Claude. Upload this repository's SKILL.md as a skill to reach every layer from one place, or download the skill of a single layer from its repository. Each skill loads only a small index and opens detailed files when a question needs them.
In Claude Code. This repository will become a plugin marketplace, so a single command installs every layer. Until then, clone the layer repositories into your skills directory.
With other AI tools or by hand. Every skill is plain markdown. Start from the SKILL.md of the layer you need.
tricot (triadic comparison of technology options) is the first domain. Since 2011 it has produced on-farm trial data on more than 30 crops in more than 25 countries, standard operating procedures in several languages and a body of peer-reviewed evidence. Environment data are available too. Daily rainfall (CHIRPS), temperature (CHIRTS, AgERA5), elevation and soil (SoilGrids) can be fetched for any point, for use with tricot trials or on their own. RHOMIS household data come next, with the Climate Action team. Further domains join as they meet the same standards.
Each layer lives in its own repository, with its own versions, license and DOI. This repository links them.
| Repository | Layer | What it holds | Status |
|---|---|---|---|
| tricot-knowledge-base | Literature | Verified cards for the open-access tricot literature, with an expert skill | Active, skill available |
| tricot-protocols | Methods | Standard operating procedures, surveys and trait metadata, in five languages | Available, skill planned |
| tricot-data | Data | Harmonized trial data in JSON, archived on Zenodo | Available, skill planned |
| environment-data | Data | Daily weather, elevation and soil properties for any point, from CHIRPS, CHIRTS, AgERA5, Copernicus DEM and SoilGrids, as agent skills | Active, skill available |
| tricot-analysis-workflows | Analysis | The team's analytical approaches, as curated use cases with decision records on methods and their limits, including how trial data link with climate, soil and other environmental data | Active, skill available |
| gosset, ClimMobTools, climatrends, chirps and ClimMob-analysis | Analysis | The R packages the workflows run on, for ranking models, ClimMob data, climate indices and rainfall, plus the workflow behind ClimMob reports | Active, packages on CRAN, report workflow on GitHub |
| template-repo-data-analysis | Analysis | A template for FAIR and reproducible analysis projects, with decision records, tests and provenance linking outputs to data, protocols and evidence | Active, skill available |
maniva
(index of the layers, one entry point)
│
┌────────────────┬──────────────┴────┬───────────────────────────┐
│ │ │ │
knowledge base protocols data analysis
what we know how we do it what we found how we reason with it
│ │
┌────────────┬───────────┬──────────────┼───────────────┐ │
│ │ │ │ │ │
on-farm breeding agronomy environmental socio-economic │
testing CHIRPS, CHIRTS, RHOMIS │
tricot soil │
available under under available under │
development development development │
│ │ │ │ │ │
└────────────┴───────────┬──────────────┴───────────────┘ │
│ │
shared keys │
what, where, when │
│ │
│ ┌───────────────┬───────────┴───┐
│ │ │ │
└────────► workflows R packages template
use cases and gosset and the reproducible
decision records ClimMob suite FAIR projects
Each layer holds one repository per domain, and repositories are named by their domain, so the tricot repositories keep their names as other domains join. A domain adds a repository to a layer only when it has that kind of content, and most domains will add data and workflows alone.
The data layer grows by domain. On-farm testing with tricot and environment data are available today. Breeding, agronomy and socio-economic data are under development, and each will live in its own repository that also works on its own. The domains connect through shared keys, what was tested, where and when. A fourth key for the participant or household is proposed. The analysis workflows hold the use cases that join the domains.
The key called what depends on a common identity for germplasm. The design, and the points open for discussion with partners, are in docs/design/germplasm-integration.md. How the suite stays readable for AI assistants as the data grow, including a proposed agent layer of trial cards and CSV files, is in docs/design/agent-access-and-scale.md.
The layers connect through shared identifiers. A paper card names the dataset its findings rest on, a dataset names the protocol that produced it, and protocols and data share trait definitions from the Crop Ontology. An assistant can follow those links from a finding to the trials behind it.
The analysis layer has three parts. The workflows say which method to use and why, the R packages run it, and the template packages a new analysis so others can rerun, trace and cite it. The template draws on the other layers too, since its provenance entries link each output to the data, protocol and evidence behind it.
- One repository per layer. Each stays the single source of truth and keeps its own citation and DOI. Nothing is copied into this repository.
- The same skeleton everywhere. Every layer has a
SKILL.md, aREADME.md, aCITATION.cffand aLICENSE. - Light skills. Skills travel between users, so they load a short index first and detailed files only on demand. Large source files stay in the repository, out of the skill.
- Verified content. Numbers are checked against their sources and carry their primary citation. Uncertainty and disagreement are stated, not smoothed over.
- History is never rewritten. Corrections are new commits logged in
CHANGELOG.md, and anyone can check our history against an independent archive. SeePROVENANCE.md. - Open by default. Content is released under open licenses, and every skill tells the user how to cite the work it draws on.
- Create the repository with the shared skeleton.
- Write a
SKILL.mdwith a short description of when to use it, instructions for the assistant, and an index of the detailed files. - Keep the skill under about 20 KB of index, with detail in files opened on demand.
- Add the repository to the table above, to the routing table in
SKILL.mdand tollms.txt.
metadata/layer-guidance.md says what each layer should document so that claims can be traced back to their sources.
maniva is pronounced ma-NEE-va. The word comes from Old Tupi mani'yba, the stem of the manioc plant. Cassava is propagated from these stem cuttings, and each one carries the potential for new growth. The crop began in South America and travelled from the Amazon to Africa and Asia, adapting to new environments on the way. maniva is meant to work the same way, so that knowledge developed in one place can take root, adapt and create value elsewhere. Cassava is also one of the crops in the trial data. The full reasoning, and the other names we considered, are in metadata/naming.md.
Rooted in biodiversity. Growing through collaboration.
To cite this documentation as a whole, use CITATION.cff.
de Sousa K, Laporte MA, Youngberg B, Cremaschi A, Chase R, van Etten J (2026). maniva. An open commons of evidence, protocols, data and analysis for agricultural research. Version 1.0.0. https://doi.org/10.5281/zenodo.24011051
This DOI always resolves to the latest version. To cite one version exactly, use its own DOI on Zenodo. Version 1.0.0 is https://doi.org/10.5281/zenodo.24011052.
When you use a specific layer, cite that repository too, and cite the original papers for any finding.
Documentation in this repository is released under CC BY 4.0, with rights held by the International Plant Genetic Resources Institute (Bioversity International). Across the suite, data are released under CC0 1.0, software under the MIT License, and documentation, protocols and methods under CC BY 4.0.