Physics IDE

A theoretical framework is a large, interlocking body of axioms, derivations, notation and manuscript. A chat window cannot hold it, and an assistant that only sees the last few messages cannot reason about it. Physics IDE puts the theory itself in the workspace — master axiom file, chapters, notes, scripts, experiments and data — and gives its AI assistants a structured, indexed view of all of it, so a question about one derivation can be answered against the whole.

What it does

• Theory profiles: one workspace per theory, with a master axiom file, structured manuscript chapters, a scratchpad, and linked tools, scripts and datasets.
• Two independent AI research lanes of configurable strength, so a strong model can check a fast one, or two lines of reasoning can run side by side on the same context.
• Project awareness built from the actual files: a generated awareness document, a theory-agnostic structural schema, and a local hybrid search index (full-text plus vector) that retrieves the relevant sections for each request.
• Gated file access: the AI is advisory by default and can be granted write access limited to the active project root when you want it to edit the framework directly.
• An AI Request Inspector and a context-load meter that show exactly what was sent, in what order, and what it cost in tokens — before and after each request.
• Manuscript search with section navigation, KaTeX rendering, and Pandoc export.
• Local AI through Ollama as a first-class provider, alongside Gemini and OpenAI: a hardware probe, a model catalog with fit and response-time estimates, a quantization selector for lighter downloads, and one-click installation.
• An agentic loop under your control: the AI can create, read, append to and edit files inside the project root, run allow-listed commands and scratch calculations with inline figures, search the index itself, and record decisions in an append-only project memory — with a git-style checkpoint taken before every edit so anything can be reverted.
• Task mode: hand a pane a multi-step goal and follow its plan with round and token budgets, Pause, Approve-each-step and Stop.
• Manuscript tools: Math Fixer (Unicode and shorthand math to clean LaTeX), master-axiom compilation from the theory sources, master-manuscript assembly, and PDF / DOCX export.
• Everything stays on the machine: SQLite indexes, local embeddings, no server of ours in the loop.

Where it stands, and a note on AI models

Version 9 is the first public release. It follows a full interactive smoke test of the file-operation, retrieval, checkpoint, task and scratch-computation subsystems on a mid-range laptop. In day-to-day research use a serious theory consumes context quickly and cloud token pricing makes sustained work expensive, so the intended fit is a local model served through Ollama: a 3B–7B instruct model runs comfortably on a 16 GB machine for routine work, with a cloud provider in the other pane for the heavy reasoning. The retrieval and prompt-assembly work is what keeps a local model viable for this workload.

Who it is for

Researchers and serious students developing or evaluating a theoretical framework — mainstream, hybrid or non-standard — who want AI assistance that is grounded in their own material rather than in whatever a chat model happens to recall.

Support

Questions, bugs and feature requests: gregorydearth@gmail.com.