Alex Station

Alex Station is a local-first AI workbench for context engineering, prompt compilation, project memory, model switching, and agent handoff.

The first shipped tool is QRPT: Quantum Recursive Prompt Tree. QRPT takes a vague intent, expands competing interpretations, scores them with a deterministic heuristic, attaches fresh evidence when available, and emits an agent-ready prompt with traceable assumptions, constraints, and unresolved context.

Abstract

Most prompt tools flatten ambiguity into a single fluent answer. QRPT keeps ambiguity inspectable. It turns intent into a tree, scores branches, builds a semantic DAG, records missing context, and finalizes only after the strongest path is visible.

vague intent
  -> interpretation tree
  -> scored branches
  -> semantic DAG
  -> evidence and assumptions
  -> validation checklist
  -> compiled handoff prompt

Repository Role

This repo owns the local interface and QRPT proof surface for the public Motherlabs spine.

• src/qrpt.ts owns the QRPT compile logic.
• src/App.tsx owns the React workbench.
• src-tauri/ owns secure local commands, OS keyring access, and provider calls.
• benchmarks/qrpt/ records structural and quality benchmark reports.
• docs/ARCHITECTURE.md maps the interface and compile graph.

Related repos:

• context-world-model: context as world-model thesis and training prototype.
• mc-v2: semantic compiler harness.
• motherforge: governed runtime-pack compiler kernel.

Stack

• Tauri v2
• React + TypeScript + Vite
• Rust command layer
• pnpm
• local JSON state in the app data directory
• OS keyring for API keys

Development

pnpm install
pnpm build
pnpm tauri dev

Browser-only UI iteration:

pnpm dev

Browser-only mode falls back to localStorage for app state and cannot securely store API keys. Use the Tauri app for API key storage and web/search provider calls.

QRPT V1

QRPT currently provides:

• project-aware prompt-tree runs
• semantic DAG output for tree paths, constraints, missing context, final artifacts, and evidence
• local and frontier model profile slots
• Brave Search and OpenAI web-search adapter plumbing
• fast deterministic, AI-assisted, and deep-compile modes
• secure API key path through Rust and the OS keyring
• tree node scoring and inspection
• auto-finalized compiled prompt
• Markdown and JSON export
• JSON import

Benchmarks

Run the QRPT structural benchmark:

pnpm benchmark:qrpt

Run the QRPT quality benchmark:

pnpm benchmark:qrpt:quality

Refresh committed baseline reports:

pnpm benchmark:qrpt -- --write
pnpm benchmark:qrpt:quality -- --write

The structural benchmark checks that vague intent compiles into a traceable tree, acyclic semantic DAG, final handoff prompt, assumptions, unresolved-context ledger, validation checklist, and evidence citations when fresh context is supplied. The quality benchmark uses a deterministic rubric by default and can use OpenAI as a live evaluator when OPENAI_API_KEY and QRPT_OPENAI_EVAL=1 are set.

Privacy Boundary

Alex Station is local-first.

API keys are not written to the repository, exported trees, frontend storage, or the station state file when running through Tauri. Provider calls that need secrets run through the Rust command layer and the OS keyring.

License

Source-available, all rights reserved. See LICENSE, NOTICE, and PROVENANCE.md.

The code and writing are public for reading, inspection, citation, and non-commercial evaluation. Commercial use, redistribution, model training on the contents, hosted-service reuse, and relicensing require written permission from Alex Roze / Motherlabs.