EvoOntology plugin
A Claude Code and Codex plugin that builds an ontology over your tables and files and refines it from the agent's work history
- 1,153 ★ on the repository
- Last commit
- MIT
Extended access
Why this level
- Reads tables, files and databases to build the ontology and writes concepts, examples and evidence from your data into the workspace
- Records the agent's trajectories and uses them to edit the ontology, so fragments of data may end up in them
Install
- Works in
In your terminal, with SkillFoxx CLI
npx skillfoxx add plugins/evoontologyDetects the agents on your machine, shows the access level and pins the version.
Other ways to install
Access and checks
How we reviewExtended accessThe tool runs code or makes network calls and only reads external data.
- Reads tables, files and databases to build the ontology and writes concepts, examples and evidence from your data into the workspace
- Records the agent's trajectories and uses them to edit the ontology, so fragments of data may end up in them
This is third-party code. Review the repository files before installing.
What it does
EvoOntology gives a data agent a dictionary of meaning: metric definitions, entity relationships and constraints that table and column names do not show. The plugin builds this layer over your tables, files and databases and opens it to the agent through the MCP tools `browse_semantics` and `resolve_semantics`. The agent requests only what the current step needs, and the whole ontology stays out of context. The plugin has three workflows. `/evoontology:build-ontology` builds the first version: candidate concepts come from the workload and are written only when the underlying data confirms them. `/evoontology:evolve-ontology` analyzes recorded agent interactions, proposes a localized patch and accepts it only when a paired comparison with the previous version shows a reproducible gain. `/evoontology:explore-ontology` opens an explorer of the content, schema and tool layers. Codex has the same three workflows as `$build-ontology`, `$evolve-ontology` and `$explore-ontology`.
Who it is for. For analysts and data engineers whose Claude Code or Codex agents write SQL and analyze reports over unfamiliar tables and misread what fields mean.
Good fit when
- An agent writes SQL over a database with dozens of tables and misreads columns, so you need a layer of metric definitions and relationships
- The agent repeats the same data exploration session after session, and what it found should be kept as verified knowledge
- Several agents must rely on one version of definitions: the ontology is versioned as `ontology_v0`, `ontology_v1` and onward
- You want to see what the agent learned about your data: `/evoontology:explore-ontology` shows concepts, mappings, constraints and evidence
- You evaluate agents on text-to-SQL, financial report research and business analytics: the repository has BIRD, DDR-10K and InsightBench adapters
Not a fit when
- You have one small table with clear column names: a semantic layer adds little
- You need a ready semantic model for a BI system with manual governance: EvoOntology builds and edits the layer automatically
- You work in Cursor or another agent without Claude Code or Codex plugin support
- Data must not be shown to the agent: the plugin hands it facts about the structure and content of your sources
Example request
Build an ontology for sales.db with the EvoOntology plugin and show in the explorer which metrics it containsLimitations
- Plugins exist only for Claude Code and Codex. Other agents can connect over MCP, but there are no ready instructions for them.
- The built-in read-only task replay works for SQLite. For other sources the agent uses its own tools and must record observations explicitly.
- The performance table is based on a four-model subset from the authors' paper and depends on the model and the task set.
- The repository is research code that accompanies a paper, and the `USAGE.md` guide sets the data boundaries.
How to disable. In Claude Code run `claude plugin uninstall evoontology@evoontology`, and in Codex remove the plugin with the Codex plugin manager. Built ontology versions sit in the project workspace and stay after removal, so delete them by hand.
FAQ
How do I install EvoOntology in Claude Code?
Run `claude plugin marketplace add ruc-datalab/EvoOntology`, then `claude plugin install evoontology@evoontology`. You do not need to clone the repository or install a pip package. After a new session starts, the commands `/evoontology:build-ontology`, `/evoontology:evolve-ontology` and `/evoontology:explore-ontology` are available.
Does EvoOntology work in Codex?
Yes, Codex has a separate plugin, `evoontology-codex`. Install it with `codex plugin marketplace add ruc-datalab/EvoOntology` and `codex plugin add evoontology-codex@evoontology`. The workflows are called `$build-ontology`, `$evolve-ontology` and `$explore-ontology`.
What is an ontology layer for a data agent?
A versioned graph of meaning: concepts, their mapping to tables and fields, constraints and evidence. The agent takes metric definitions and entity relationships from it and does not guess them from column names.
How does EvoOntology decide that an ontology update helps?
A new version publishes after a paired check. The candidate and the parent get the same data, the same agent, the same decoding settings and the same interaction budget. Without a reproducible gain the parent stays, and the result feeds the next round.
What results does EvoOntology show on benchmarks?
According to the authors' table on a four-model subset, BIRD execution accuracy rose from 63.6 to 72.4 and the DDR-Bench trajectory-wise score from 69.5 to 89.5. The comparison was a ReAct agent without an ontology against an agent with the self-evolving ontology.
Discussion
No questions or reviews yet.
Related
PluginsSemantica plugin: 17 skills and 3 agents for knowledge graphs, reasoning and data provenance in Claude Code, Cursor and Codex.
Open Ontologies
open-ontologies
Planning changes to OWL ontologies: impact assessment, rollback and verifiable evidence for a reviewer. A single Rust binary for a narrow audience
The UModel plugin gives an agent three skills to read entities and topology from a graph and trace root causes locally.
A Claude Code plugin and MCP server: over 1000 scientific tools for biology, chemistry and medicine, data-check commands and routing skills
A CLI and MCP server that extracts structured knowledge from documents: graphs, hypergraphs, and fact search with sources
A plugin for Claude Code, Codex and Copilot that maps ML training code into a diagram linked to source lines
Comparison with similar tools
Cards that share the main search query, and how they differ.
| Tool | Official | Authentication | Repository stars | Main difference |
|---|---|---|---|---|
| EvoOntology (this card) | no | 1,153 | ||
| Open Ontologies | no | not required | 2,829 | plans and verifies changes to existing OWL ontologies; EvoOntology builds an ontology from your data and refines it from the agent's history |
| MCP Knowledge Graph | no | not required | 888 | persistent agent memory in a local knowledge graph; EvoOntology ties concepts to tables and verifies them against the data |
README in short
The README opens with the gap between agents and data and a summary of the project's principles. It then shows the three ontology layers with explorer screenshots, a five-step lifecycle, plugin install and a results table on three benchmarks. It ends with the repository layout, documentation and a link to the paper.