Give every agent the same world to reason from

Penumbra gives agents a governed, source-backed model of your domain, available through MCP in the harness where work already happens.

Codex, Claude Code, Cursor, and similar tools are agent harnesses: they provide a capable model, a workspace, and tools for the work in front of it. What they do not provide is a durable, governed understanding of your domain.

Every new session begins with a reconstruction: a prompt, a handful of files, search results, and whatever context happens to fit. The pieces may be accurate while the overall meaning still drifts. A customer, a risk, or a decision can mean something different to each agent, with no durable record of why.

Penumbra provides that cognitive substrate: a shared model of your domain that persists across agents and sessions. Sources remain attached to the knowledge derived from them. Shapes express the types, relationships, and standards that matter. Consequential changes can be inspected before they become shared truth.

Connect your agent

Add Penumbra to Codex, Claude, Cursor, or another MCP-capable harness.

Run the first workflow

Save a source, retrieve it, and verify the exact wording in a few minutes.

Why coherence beats more context

Longer context windows and better retrieval are useful, but they solve a different problem. They help an agent see more material in one moment. They do not decide what that material means, preserve the provenance of an interpretation, or keep the interpretation consistent across tools and time.

Penumbra holds the layer those systems leave implicit: the domain model, source-backed knowledge, review state, and memory agents use to orient themselves. Different agents can then enter the same project with the same definitions, evidence, and operating boundaries.

One shared substrate, three capabilities

Shape the domain

Make your way of understanding a domain explicit as types, fields, relationships, and constraints.

Work from evidence

Let agents find sources, retrieve structured knowledge, and stage reviewed changes against a live project.

Carry what matters

Preserve decisions, preferences, and lessons across sessions without confusing memory with source-backed truth.

Shapes, Runtime, and Memory are separate MCP servers because they govern different kinds of work. Together, they let an agent understand the project it has entered, act within its boundaries, and return reviewed knowledge or proposed changes to the same project.

Use it in the harness you already have

MCP makes the substrate available without forcing the work into a new chat surface. Codex, Claude Code, Claude.ai, Google Antigravity, Cursor, and other harnesses keep providing the agent, workspace, and interaction model. The Penumbra MCPs give that agent access to the persistent project.

Connect the MCPs in the harness where you already work, then describe the outcome in plain language. The agent can inspect the active project, discover the Shapes it should use, retrieve the underlying evidence, and show you proposed changes before applying them. Moving between harnesses does not require rebuilding the project’s context from scratch.

See the system at work

Turn documents into governed knowledge

Preserve the original material, extract it through a Shape, and verify the result against the source.

Build a team brain

Give people and agents shared project knowledge plus decisions and lessons that persist across sessions.

Answer from stored evidence

Gather records and source passages, then make gaps and inferences explicit before acting.

Run an expert interview

Preserve the conversation and review the findings before they become part of the project.

Penumbra is in preview for design partners and developer testers. Request access at [e-mail].

Feedback

If a workflow is missing, confusing, or wrong, email [e-mail].