AI Software Development Platform

Build software with an AI that remembers how you build.

Lumen is a local-first AI software development platform. The daemon runs on your machine, keeps a persistent brain across sessions, and routes every request through the right model, role, and skills. Reach it from any browser via Cloudflare Tunnel.

Operator, not chat

Lumen routes between the Anthropic SDK and the Claude Code CLI per task, picks the model, shapes the context, and captures every outcome. The chat window is one surface, not the whole product.

Local-first by design

SQLite databases, vector indexes, and an encrypted vault live in your home directory. Your code, secrets, and history never leave the machine unless you decide they should.

Persistent context

A per-workspace brain stores directives, patterns, and episodic memory across sessions. Hybrid retrieval (BM25 plus vector, fused with reciprocal rank) surfaces the right context on every turn.

Phone-installable

The dashboard is a PWA. Add it to your iOS or Android home screen for a standalone app experience, served from your laptop over a Cloudflare Tunnel.

How it works

Lumen is built around the operator pattern: every request is routed, shaped, executed, and captured. Four layers do the work.

Router
Selects the model and executor. Routes between the Anthropic SDK (for fast structured replies) and the Claude Code CLI (for filesystem and shell tasks) based on intent.
Context Optimizer
Shapes the prompt. Composes the active role, installed skills, workspace directives, and brain-retrieved patterns into the right system prompt for this turn.
Executor
Runs the call. Streams tokens back over Server-Sent Events with 15-second heartbeats so mobile middleboxes do not drop the connection mid-thought.
Outcome Capture
Learns from every turn. Persists token usage, latency, model, stop reason, and pattern matches so the brain gets better with every session.

Specs

Built for builders. No tracking, no telemetry, no surprise cloud calls.