The substrate
agents run on.
Five primitives turn ad-hoc agent usage into a real platform: workspaces, ingestion, typed primitives, multi-agent threads, and an MCP surface every tool can speak to.
Workspaces & Multi-Agent Runtime
A CortexGraph workspace is the unit of company context — one shared graph that every human and agent operates against. The runtime is harness agnostic: Cursor, ChatGPT, Claude, and your own custom agents are all first-class actors with scoped permissions, attribution, and shared state.
Knowledge Ingestion
Ingestion pipelines pull docs, decisions, repos, customer threads, and runbooks into immutable raw sources, then compile them into a maintained markdown graph with provenance back to every original document. Knowledge is captured once and becomes reusable context for every agent that joins the workspace.
Primitives, Threads & Long-Horizon Runs
Define the typed primitives your business actually thinks in — decisions, accounts, projects, runbooks, incidents — and let agents read and write them directly. Threads, runs, and checkpoints turn multi-day, multi-agent work into durable state instead of disposable chat history. Any agent can pick up where another left off.
MCP as the Plug & Play Surface
One MCP endpoint, OAuth-secured, exposes the entire workspace — graph, primitives, threads, runs, decisions. Cursor, ChatGPT, Claude, and any custom agent connect in seconds and immediately have the same context as every other agent in the company. No SDK, no per-tool onboarding doc, no glue code.