Cascade

In development

A local-first AI agent runtime

One CLI and daemon for agents, context, model routing, plugins and storage that runs the same from a laptop to a server. Cascade began as a handful of scripts to work around the walls I kept hitting in Claude Code: one account running out mid-task, token spend on routine operations, and context that reset between sessions.

Source on GitHub Read the design notes

MIT licensed. Version 2 is in early development; no v2 release yet.

Where it stands

Version 2 is a clean-sheet rewrite in Go: a single static binary with no CGO and no external build dependencies, for macOS, Linux and Windows. The complete v1 implementation, released through v1.16, is archived read-only atacamarata/cascade-v1. Cascade is the open-source version of patterns that go further inside nSelf and nClaw, but the core holds up on its own.

Build from source

# Go 1.26.2 or newer
git clone https://github.com/acamarata/cascade.git
cd cascade
go build ./...

What it does

Model routing across accounts

Conductor picks the cheapest capable model for each kind of task and keeps a live ledger of which account still has quota, so work continues when one account is throttled.

Persistent local daemon

A background daemon owns long-running sessions, fleet state and scheduling, and survives restarts.

Context, retrieval and memory

Hybrid retrieval over the project's own files and notes, with memory that carries across sessions instead of starting from zero.

Agents, plugins and integration

Agents and workflows run from one CLI, with a local API and MCP server for other tools, and a plugin boundary for what does not belong in core.

Writing

Open source, MIT licensed

Issues and contributions welcome while v2 takes shape.

github.com/acamarata/cascade →