get-physics-done

What it does

Get Physics Done (GPD) is an open-source AI agent orchestration framework designed specifically for structured physics research. Rather than a single-prompt LLM interaction, GPD decomposes a research problem into four phases — Formulate, Plan, Execute, and Verify — and coordinates specialist sub-agents through each phase. The goal is to go from a vague hypothesis to a formatted arXiv-ready LaTeX manuscript.

The tool is built on the Model Context Protocol (MCP) and runs as a layer on top of an existing LLM runtime (Claude Code, Gemini CLI, Codex, or OpenCode), meaning it does not supply its own model but orchestrates whichever model the researcher already has access to.

Key capabilities

• Four-phase structured research workflow: Formulate → Plan → Execute → Verify
• Physics verification layer: automated checks for dimensional consistency, limiting cases, symmetry constraints, conservation laws, and numerical stability
• Multi-agent teams: separate specialist agents for high-level theory, numerical simulation/coding, and LaTeX manuscript preparation
• Structured research memory: persists context across long-horizon projects (days or weeks of compute)
• End-to-end output: produces arXiv-ready .tex manuscripts, Python verification scripts, and figures
• Three operating modes: research copilot, manuscript peer reviewer, and autonomous research autopilot
• arXiv integration via arxiv-mcp-server for literature retrieval
• 61 CLI commands covering the full research lifecycle

Fit for Sheffield researchers

Most relevant to researchers in Physics, Applied Mathematics, and computational Engineering at Sheffield. The structured four-phase workflow with built-in sanity-checking addresses a real pain point: researchers under time pressure often skip dimensional and limiting-case checks that an AI agent can run automatically and tirelessly.

The tool requires an existing paid LLM runtime subscription. Researchers using Claude Code or Gemini CLI already have a compatible backend. Because GPD sends research content to a third-party LLM API, unpublished work, sensitive simulation data, or commercially sensitive material should not be used without reviewing the data handling policy of the chosen backend runtime.

GPD is best suited to long-horizon, structured computational and theoretical physics projects. Lighter tasks (quick literature searches, single derivations) may not justify the setup overhead.

Licensing and access

Apache 2.0 licence. Install via npx -y get-physics-done (requires Node.js) or pip install get-physics-done (Python 3.11+). No institutional licensing arrangements exist at this time; the tool itself is free but depends on a paid LLM API at runtime.

Trust and governance notes

GPD does not handle data itself — all content is passed to the chosen LLM runtime (Claude Code, Gemini CLI, Codex, or OpenCode). Data handling, GDPR compliance, and data residency are therefore determined by the selected backend, not by GPD. Researchers must review the data policy of their chosen runtime before using research data.

No formal privacy policy is published by Physical Superintelligence PBC at the time of this review. The tool is too new (released March 2026) for independent governance assessment. SSO is not supported; authentication is managed by the underlying LLM platform.

Observed limitations

• Publicly released March 2026 — too new for independent benchmarking or failure-mode analysis
• Does not generate novel physical intuition; structured verification is rigorous but bounded by known constraints (dimensional analysis, conservation laws); genuinely unprecedented problems may not be well served
• Dependent on a paid third-party LLM runtime — cannot be run fully on-premises
• Data governance is delegated to the LLM backend; no vendor GDPR statement available
• Windows installation issues reported in community discussions
• Community skepticism (noted in video comments) that the agent can follow physics rules convincingly but may struggle with truly novel “out-of-the-box” conceptual leaps
• Real-world validation of the human-in-the-loop AI physics workflow is provided by the Schwartz/Anthropic vibe-physics case study (arXiv 2601.02484): using Claude Opus 4.5 + Claude Code, a Harvard QCD expert produced a full research paper in ~2 weeks (vs typical 1–2 years), but at the cost of ~50–60 hours of expert oversight, 51,248 messages, and ~36M tokens. The AI was characterised as a “second-year graduate student” — competent at well-defined calculations but prone to fabricating results, requiring constant domain-expert verification. This is precisely the oversight model GPD’s Verify phase is designed to support.

Related tools

chemcrow · ag2-autogen · google-co-scientist · paperqa2

Sources and evidence

• GPD GitHub repository: https://github.com/psi-oss/get-physics-done (accessed 2026-03-24)
• GPD PyPI package: https://pypi.org/project/get-physics-done/ (accessed 2026-03-24)
• Physical Superintelligence announcement (Substack): https://theinnermostloop.substack.com/p/physical-superintelligence (accessed 2026-03-24)
• Apache 2.0 License file: https://github.com/psi-oss/get-physics-done/blob/main/LICENSE (accessed 2026-03-24)
• Anthropic “Vibe Physics” case study: https://www.anthropic.com/research/vibe-physics (2026) — documents Harvard professor Matthew Schwartz’s use of Claude Opus 4.5 + Claude Code to produce a QCD paper (C-parameter Sudakov shoulder) from start to arXiv in ~2 weeks
• arXiv 2601.02484 — Schwartz et al.: the QCD paper produced via the vibe-physics workflow; provides primary evidence for capability and limitation claims above