Skip to main content
Penumbra takes what your business knows, the data in your systems and the way your team thinks about it, and turns it into governed, queryable structure your agents can reason over and act on. You ship agents that behave like they understand your domain, and that you can trust in production. Pull from the systems you already run, model your domain with shapes, reason over one governed store, and wield it from agents, copilots, pipelines, and apps.

​
Why coherence beats more context

An LLM call is stateless and a vector search is lossy: you get retrieval, not coherence. Spread your meaning across a vector database, a schema registry, prompt strings, and a graph you patch by hand, and the same concept ends up meaning three different things in three systems. Your agents inherit the drift. Penumbra holds one high-fidelity representation of your domain, governed and provenanced, so every agent and every call reasons over the same meaning. Coherence and fidelity are what make agent behavior trustworthy. Not more tokens.

​
How Penumbra works

Six capabilities cover the lifecycle, from defining your domain to acting on it.
  • Define your domain — shapes. Typed frameworks that make your domain explicit, authored in the Shapes Workbench. Everything else reads from them.
  • Ingest and extract — sources. Bring documents and data in, and coerce them into typed structure through a shape. The same source read through different shapes surfaces different meaning.
  • Govern every change — semantic git. Every write stages, plans, applies, and reverts. Version control for meaning, with lineage preserved.
  • Retrieve and reason — search. Hybrid, semantic, and lexical retrieval over entities and the sources behind them.
  • Check before acting — decision quality. Ask whether a region of the graph is fit to act on, for a specific purpose. A verdict with findings, never a confidence score.
  • Interop in and out — import and export. Import OWL and RDF, and compile your domain back out to OWL, RDF, SHACL, JSON, YAML, and Markdown. Define once, project everywhere.
You build against two surfaces: the SDK (@penumbra-systems/platform), typed TypeScript access most builders use, and the API, the REST endpoints the SDK calls at https://pnbr.io/v1.
Penumbra is in preview for design partners and developer testers. Request a key at shep@getpenumbra.ai.

​
Start building

Design your shapes

Author your domain model in the Shapes Workbench, the design-time MCP.

Install the SDK

Add the client, set your key, and connect in about two minutes.

Quickstart

Read your ontology, search, capture, recall memory, and check decision quality.

Ingest a document

Upload, extract through a shape, and coerce material into the graph.

​
Feedback

This is an early surface and we want to hear from you. If something is missing, confusing, or wrong, or you have an opinion about where this should go, email shep@getpenumbra.ai.