A fact store for AI agents, on Postgres: registered attributes, every write citing its evidence.
Runs on your machine from its package; the steps are below.
At most: creates or changes data: records, files, settings.
Odel charges nothing per call. The publisher may charge for its own service.
Runs locally over stdio
This server isn't hosted — your MCP client launches it from a package registry. Use one of the commands below, or drop the config into your client (e.g. Claude Desktop).
Run
uvx factstore
MCP client config
{
"mcpServers": {
"factstore": {
"command": "uvx",
"args": [
"factstore"
]
}
}
}A fact store for AI agents, on Postgres. Agents record what a company's systems and documents say as facts: a value for a registered attribute, in a transaction stamped with who wrote it and the document it came from. Nothing is overwritten, so the store can say what it knew on any date.
Skills that ship with it index a company's systems from their exports, and read documents and mail into facts. They leave people's personal details in the source. MIT-licensed.
pip install factstore
In Claude Code, the plugin brings the MCP server and the skills:
/plugin marketplace add Victor-EU/factstore
/plugin install factstore@factstore
factstore/README.md walks through Postgres, a first store and connecting an agent.
factstore/ | The kernel, the MCP server, the Python SDK and the CLI |
packages/ | Vocabulary packages: core, and ecom-ops and ecom-index for an online brand |
factstore-skills/ | The catalogue, ontology and ingestion skills |
evals/ | What it was tested on: a synthetic brand, and five public datasets from real companies |
factstore-design.md | The design, and what testing changed in it |
factstore-build-plan.md | The milestones, and where each stands |
0.1.0, alpha. The kernel's code held unchanged through every test, up to 3.2 million facts in one store. The skills passed every round on public data: product lists, a million order lines, a marketplace's duplicate customers, 100 suppliers' forms and a trader's mailbox. No company has run it on its own data yet.