Cascade turns one big ask into a dependency graph of small, bounded subtasks,
then fans them out across parallel Claude Code or Codex agents.
Watch every node light up live in your browser as it runs.
~/your-project · zsh
➤ ~ $ claude mcp add --scope user cascade -- npx -y @pepps233/cascade
✓ added stdio MCP server cascade to user config
type /cascade in Claude Code to start a run
GRAPH ENGINEERING
Nodes have boundaries. Edges carry meaning.
A linear agent workflow serializes work that never needed to wait. Cascade models
the task as a graph instead: each node is a bounded unit of work with explicit
success criteria, each edge is typed and carries exactly the context that needs
to cross, nothing more.
WHY IT WORKS
Every node earns its place in the graph.
Most tasks contain work that could run concurrently, evidence that would overload
a single context window, and checkpoints that need precise placement. A graph
handles all three.
RUNS IN PARALLEL
Independent branches execute concurrently because nothing in the graph forces them to wait on each other.
SCOPED CONTEXT
Every node sees a bounded slice of the problem. Edges carry only the evidence that actually needs to cross.
CONTAINED FAILURE
A node has explicit success criteria, so it can be retried on its own without rerunning the whole graph.
LIVE BROWSER VIEW
Local viewer opens the moment a graph is created. Watch nodes go ready → running → done in real time.
CLAUDE + CODEX
Holds no model access of its own. It shells out to the CLIs already configured on your machine.
STDIO, LOCAL
Runs as a local MCP server over stdio. Your code and your keys never leave your machine.
GET STARTED
One command. Every core, working.
install
➤ ~ $ claude mcp add --scope user cascade -- npx -y @pepps233/cascade
then type /cascade inside Claude Code to break your first task into a graph.