Origin Pro MCP

Origin Pro MCP Server

An MCP (Model Context Protocol) server that enables AI assistants like Claude to control OriginLab Origin Pro via COM automation. All operations are reflected in Origin's GUI in real-time — you watch as the AI creates worksheets, plots graphs, and styles figures.

What Can It Do?

• Worksheet Management — Create workbooks, read/write data, import CSV/Excel, export CSV, column formulas, sort, transpose
• Matrices & 3D — Matrix books, XYZ gridding, 3D surface/scatter, contour, heatmap, image plots
• Graph Creation — Scatter, line, line+symbol, column, bar, area, pie, histogram, contour plots
• Graph Layers & Axes — Log scales, dual Y axis, panels, reference lines, text annotations
• Plot Styling — Colors, symbols, line width, publication-ready formatting in one call
• Analysis — Curve fitting, FFT, smoothing, integration, differentiation, interpolation, peak finding
• Statistics — Descriptive stats, two-sample t-test, frequency counts (via the stats and transform tools)
• Project Management — New/save/load projects, export all graphs (with pixel-size control)
• LabTalk Scripting — Direct LabTalk execution with destructive/file-overwrite commands blocked

This MCP server is intentionally Windows-runtime-only. The AI agent or MCP client can run from Windows or WSL, but the MCP server process that talks to Origin must be launched with Windows Python and pywin32. WSL/Linux can edit the project and run non-COM unit tests, but cannot directly control Origin COM.

Quick Start

Agentic Setup (read this if you are an AI agent configuring this server)

Copy-paste rules to self-configure the origin-pro MCP with zero human steps:

1. Environment requirement. This server only works on Windows with a licensed Origin/OriginPro install and Windows Python with pywin32. COM automation cannot run from WSL or Linux Python. If the agent runs in WSL, it must launch the Windows executable (uvx.exe / Windows origin-pro-mcp), never WSL Python.
2. Preferred install path. Use uvx (zero setup) or pip install origin-pro-mcp. Both install to a Windows-local location and "just work":

``json { "mcpServers": { "origin-pro": { "command": "uvx", "args": ["--quiet", "origin-pro-mcp"] } } } ` From WSL, call uvx.exe instead of uvx` (see Option A below).

1. Critical UNC rule. If you point at a local checkout from WSL, the repo MUST live under a Windows drive (e.g. C:\Users\YOU\Origin-Pro-MCP). Never reference it through a \\wsl.localhost\... UNC path: the background daemon (python -m origin_pro_mcp.daemon) cannot resolve a -m module or a working directory over UNC, so it fails silently (Origin never opens). uvx/pip install avoid this entirely.
2. Discover skills first. After connecting, call the list_skills tool to see bundled skills (e.g. publication-figure) and when to use each, then get_skill(name) to load the full step-by-step instructions before producing figures/plots/analysis. No file copying is needed — skills ship inside the package and are exposed by the server.

1. Prerequisites

• Windows with a licensed Origin/OriginPro installation that exposes the Automation Server
• Python 3.10+ (Windows Python, not WSL)

Tested environment: Origin Pro 2020. Other Origin/OriginPro versions may work if they expose compatible COM Automation Server and LabTalk behavior, but they are not verified by this project yet.

2. Install & Configure

Option A: uvx (recommended — zero setup)

No manual install needed. The MCP client launches the server for you. Just add this to your Claude Code MCP settings when Claude Code is running on Windows:

{
  "mcpServers": {
    "origin-pro": {
      "command": "uvx",
      "args": ["--quiet", "origin-pro-mcp"]
    }
  }
}

uvx automatically downloads and runs the server in an isolated environment. Nothing else to install. The --quiet flag keeps first-run dependency messages out of your MCP client logs.

If Claude Code or another MCP client is running inside WSL, launch the same Windows server by calling Windows uvx.exe directly:

{
  "mcpServers": {
    "origin-pro": {
      "command": "uvx.exe",
      "args": ["--quiet", "origin-pro-mcp"]
    }
  }
}

For a local checkout before publishing/installing, point Windows uvx at the Windows path of the repo:

{
  "mcpServers": {
    "origin-pro": {
      "command": "uvx.exe",
      "args": ["--quiet", "--refresh", "--from", "D:\\04.Agent OS\\Origin-Pro-MCP", "origin-pro-mcp"]
    }
  }
}

Keep the command and args as separate JSON array entries. That avoids quoting problems when a Windows path contains spaces. If WSL cannot find uvx.exe, set command to the full WSL path for the Windows executable, for example /mnt/c/Users/YOU/.local/bin/uvx.exe.

> WSL users — the package must live on a Windows-local path, not a \\wsl.localhost\... (UNC) path. The server runs a background daemon (python -m origin_pro_mcp.daemon); Windows cannot resolve a -m module or use a working directory over a UNC path, so launching it from a WSL-filesystem checkout fails silently (Origin never opens and a console window flashes). uvx/pip install already install to a Windows-local location, so they are unaffected. If you point at a local checkout, clone it under a Windows drive (e.g. C:\Users\YOU\Origin-Pro-MCP) and reference that path — do not point the MCP config at a repo inside the WSL filesystem.

Option B: pip install from PyPI

pip install origin-pro-mcp

Then configure Claude Code:

{
  "mcpServers": {
    "origin-pro": {
      "command": "origin-pro-mcp"
    }
  }
}

Option C: Clone and run directly

git clone https://github.com/youngminsw/Origin-Pro-MCP.git
cd Origin-Pro-MCP
pip install -e .

{
  "mcpServers": {
    "origin-pro": {
      "command": "origin-pro-mcp"
    }
  }
}

> Note: If Claude Code runs in WSL, make sure the uvx or python command points to your Windows Python, not WSL Python. Origin COM only works from Windows.

3. Origin Startup

You do not need to start Python or Origin manually. The MCP client starts origin-pro-mcp, which launches an isolated Origin instance for the session. By default the Origin window is visible, so you watch the agent create worksheets and plot graphs in real time.

Display mode — set the ORIGIN_PRO_MCP_VISIBLE environment variable in your MCP config:

| Value | Mode | Use for | | :--- | :--- | :--- | | 1 (default) | Visible — Origin window shown | Watching the agent work interactively | | 0 | Invisible — Origin runs hidden | Headless/batch runs, many concurrent agents, no window pop-ups |

{ "mcpServers": { "origin-pro": {
  "command": "origin-pro-mcp",
  "env": { "ORIGIN_PRO_MCP_VISIBLE": "0" }
} } }

Reliability & recovery (advanced env vars)

The background daemon runs one isolated Origin instance per session. These environment variables harden it against a wedged Origin COM call (a synchronous operation that never returns) and against destructive mistakes. Some default off; the wedge/data-safety ones now default on because they are safe (see each row). Set any to off to opt out.

| Variable | Default | Effect | | :--- | :--- | :--- | | ORIGIN_PRO_MCP_DISPATCH_TIMEOUT | 90 | Soft budget (seconds) for each tool dispatch. If Origin doesn't respond within it, a persistent per-session watchdog is polled for the modal dialog that is blocking it: when one is found, the error names its exact title and says whether it was already auto-dismissed (retry the call) or is waiting for you to close it by hand; if none is found it falls back to the generic "most likely a modal dialog" message. NOTHING is killed at this stage. Set off/0 to disable. | | ORIGIN_PRO_MCP_DISPATCH_KILL_GRACE | 90 | Grace (seconds) AFTER the soft warning before Origin is force-killed as a last resort (so a wedged session can never permanently hold a pool slot). Total time to a hard reset = timeout + grace (default 180s). Set off/0 for no warning phase (legacy: force-kill straight at the soft budget). | | ORIGIN_PRO_MCP_DIALOG_AUTODISMISS | on | Every session runs a persistent watchdog that polls (~2s) for modal dialogs owned by its Origin process and records their titles. By default it also auto-dismisses (closes) each one it finds, so a startup or mid-session dialog no longer freezes the session. Set to 0/off/false/no to keep detection and reporting (dispatch-timeout errors still name the dialog) without the daemon closing it — you then close it by hand in the Origin window. | | ORIGIN_PRO_MCP_AUTOSAVE | on | Save the project in place (its own file, same name — like pressing Save) before a destructive op (delete graph/plot, column deletion, project load/new, overwriting a populated sheet, or a confirmed destructive run_labtalk), so a bad edit is recoverable by reloading. It never writes a differently-named copy, and never overwrites a real file with an empty/blanked project (N5-safe). Only saves a project that already has a file on disk. Set off to disable autosave entirely. | | ORIGIN_PRO_MCP_AUTOSAVE_INTERVAL | 300 | Also save the project in place every N seconds (proactive autosave), not just before destructive ops. Applies to agent-isolated sessions with a saved project; the Origin you ATTACH to is left to you. off/0 disables periodic autosave (preflight still runs). | | ORIGIN_PRO_MCP_AUTOSAVE_REQUIRED | 1 | When autosave is on and a required preflight in-place save fails, the destructive op is not run and an error is returned. Set 0 to proceed without saving. | | ORIGIN_PRO_MCP_REAP_CLOSE | off | Session lifecycle: by default a session ending gracefully (idle / client disconnect) is detached — the session's worker thread stops but your Origin window is left exactly as it was (original save path and unsaved edits intact) so you keep the project. Set 1 to restore the old save-a-recovery-copy-and-close behavior. (A wedged session's Origin is still force-killed — the only way to free a worker stuck in a synchronous COM call.) | | ORIGIN_PRO_MCP_SWEEP_ORPHANS | off | By default a restarting daemon does not kill leftover Origin windows (so a restart never destroys a project you kept open). Set 1 to have startup reclaim leftover Origins (orphan cleanup, at the cost of closing kept windows). | | ORIGIN_PRO_MCP_NO_SPAWN | off | Set 1 to stop the shim from auto-respawning the daemon. Use it to shut the daemon down from the process manager and keep it stopped — tool calls then return a clear "daemon not running" error instead of relaunching it. | | ORIGIN_PRO_MCP_ATTACH | off | Set 1 so this session attaches to the Origin you already have open (the shared ApplicationSI instance) instead of spawning a fresh isolated one — the agent then works on your currently-open project. Only one session can attach (a second falls back to an isolated instance); other agents keep their own isolated Origins. The attached instance is never force-killed by the daemon (it's yours). |

Per-call override: run_labtalk(script, confirm=True, timeout=120) bounds that one call even when a longer/shorter budget than ORIGIN_PRO_MCP_DISPATCH_TIMEOUT is needed (and works even when the timeout is off).

Rollback: unset any of these (or set the timeout to off) to return to the prior behavior — no code change or redeploy required.

Session lifecycle & restarts

Each MCP client process gets its own daemon session (its own isolated Origin). Because sessions and the daemon can restart independently, the daemon keeps a small ledger sidecar (sessions.json, next to the private lockfile) recording each session's last Origin PID and project path. When a new session starts, the daemon reads that ledger once and, on the first successful tool response, piggybacks a short one-time [origin-mcp] notice telling the agent what happened — so it continues the work instead of silently rebuilding into an empty window. What you may see:

• Your MCP client restarted (new session). Your previous Origin window was

detached, not closed (see ORIGIN_PRO_MCP_REAP_CLOSE): the notice says it is still open with your project and to save/close it in the GUI before reloading, or just work in the fresh instance.

• The daemon restarted and your old Origin is gone. The notice says your

project is not loaded and to reopen it with load_project.

• Ghost windows. Leftover Origins from earlier sessions are preserved by

default, so they can accumulate. The notice summarizes how many are still open; close them in the GUI once saved, or set ORIGIN_PRO_MCP_SWEEP_ORPHANS=1 so a restarting daemon reclaims them.

• Attach (ORIGIN_PRO_MCP_ATTACH=1). If you got the user's open Origin, the

notice reminds you that autosave and force-recovery are disabled there — save explicitly and avoid destructive ops. If another session already holds the single attach slot, the notice says you got an isolated Origin instead.

Separately, load_project appends a one-line collision warning to its result when the ledger shows another live Origin still holding the same project file (saving from both would clobber it — close the other first). The notice and the warning are advisory strings only; they never block a call, and a missing/corrupt ledger is treated as empty.

4. Use It

Just ask Claude to work with Origin:

"Create a scatter plot from this data: x=[1,2,3,4,5], y=[2.1,4.0,5.9,8.1,10.0]"
"Apply publication styling to Fig1 with axis labels Temperature (K) and Absorbance (a.u.)"
"Fit a Gaussian to the data in Book1"
"Export all graphs to C:\Users\me\figures\"

File paths can be Windows style (C:\Users\me\fig.png) or WSL style (/mnt/c/Users/me/fig.png) — the server converts WSL paths automatically, so agents running in WSL can pass their native paths.

Agent Location vs Server Runtime

The agent does not have to run on Windows. These setups are valid:

• Windows agent -> Windows origin-pro-mcp server -> Origin Pro
• WSL agent -> Windows origin-pro-mcp server -> Origin Pro

The unsupported setup is WSL/Linux origin-pro-mcp server -> Origin Pro, because COM is a Windows API.

Version Support

This project is currently verified only with Origin Pro 2020. The implementation uses Origin's COM Automation Server and LabTalk, which exist across multiple Origin releases, so other versions may work. Treat them as unverified until someone runs the test suite and a real graph/export smoke test on that version.

Direct LabTalk Safety

The run_labtalk tool is available by default for styling, analysis, graph tweaks, and other advanced Origin operations. It blocks common destructive or file-writing LabTalk commands such as project reset, delete, save/open, file dialogs, external script execution, and graph export. Use the typed tools for saving, loading, importing, and exporting.

This is an accident-prevention guard, not a security sandbox for untrusted code.

LabTalk Gotchas (Origin 2020, styling-report fixes)

• One flag per set call. set -c color(255,0,0) -cf color(255,0,0);

(combining flags in ONE command) silently wipes the plot to black; the same applies to -k/-kf/-z combined (can blank the symbol). Send each flag as its own set -flag val; call.

• Tick-label offset is layer..label.offsetV / .offsetH (vertical for

the x axis, horizontal for y), in % of the tick-label font size, positive = toward the axis. The plausible-looking names .offset, .voff, .hoff, .offsetX/.offsetY, .xOffset/.yOffset are all silent no-ops (they "read back" a value but never move the labels). Use `set_t

…

Source & license

This open-source MCP server is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: youngminsw
• Source: youngminsw/Origin-Pro-MCP
• License: MIT

Install and usage instructions live in the source repository linked above.