Skip to content
Kjean13 edited this page Nov 19, 2026 · 2 revisions

Home

cvi-mcp is a read-only MCP server for LabWindows/CVI projects. It indexes the C sources into SQLite outside the project and exposes thirty-eight tools, so an agent can answer questions about a codebase it has never seen.

It exists for one situation: you open an unfamiliar CVI project and need to know what calls what, which button runs which callback, and what changed since the last release, without reading fifty files by hand.

Three pages: this one to get running, Tools for the exact parameters and returns, Internals for how it works, what it misses and what to do when it misbehaves. En français: Accueil.

Install

Python 3.10 or newer. Static analysis runs on Windows and Linux; detecting a CVI installation and running build_check need Windows.

Windows:

py -m venv .venv
.venv\Scripts\activate
pip install -e .
cvi-mcp --check

Linux:

python3 -m venv .venv
source .venv/bin/activate
pip install -e .
cvi-mcp --check

cvi-mcp --check reports what the machine actually offers rather than what the documentation assumes: detected CVI installations whatever their year, SDK header folders, build executable, active settings, index state and warnings. It exits non-zero when a warning is present, so it also works in a script. Run it before anything else.

Connect a client

Claude Code

Project scope, .mcp.json at the repository root:

{
  "mcpServers": {
    "cvi": {
      "command": "cvi-mcp",
      "env": {
        "CVI_MCP_PROJECT": "C:\\projects\\app",
        "CVI_MCP_ALLOW_BODY": "1"
      }
    }
  }
}

Or from the CLI. Options come before the name and -- separates the server command:

claude mcp add --scope user --env CVI_MCP_PROJECT=C:\projects\app --transport stdio cvi -- cvi-mcp

Scopes: local is the default and private to the project, project is shared through .mcp.json, user applies everywhere.

OpenCode

opencode.json in the project, or ~/.config/opencode/opencode.json. Three differences from the shape above, and all three fail quietly if you get them wrong: servers sit under mcp.servers, the command is an array, and the variables block is called environment.

{
  "$schema": "https://opencode.ai/config.json",
  "mcp": {
    "servers": {
      "cvi": {
        "type": "local",
        "command": ["cvi-mcp"],
        "environment": {
          "CVI_MCP_PROJECT": "C:\\projects\\app",
          "CVI_MCP_ALLOW_BODY": "1"
        }
      }
    }
  }
}

Older versions put the server directly under mcp and use enabled instead of disabled. Check with opencode mcp list.

Claude Desktop

{
  "mcpServers": {
    "cvi": {
      "command": "cvi-mcp",
      "env": { "CVI_MCP_PROJECT": "C:\\projects\\app" }
    }
  }
}

If cvi-mcp is not on the client's PATH, which is often not the PATH of your shell, give the full path to the script or use "command": "python" with "args": ["-m", "cvi_mcp.server"].

Configuration

Variable Default Role
CVI_MCP_PROJECT . project folder, or directly its .prj / .cws
CVI_MCP_PROJECTS empty named projects for use_project: App=C:\\a;Other=D:\\b
CVI_MCP_DB ~/.cvi-mcp/index-<hash>.db SQLite index, one per project
CVI_MCP_SDK_DB ~/.cvi-mcp/sdk.db SDK index, shared by all projects
CVI_MCP_AUDIT_LOG ~/.cvi-mcp/audit.jsonl call log
CVI_MCP_EXCLUDE empty excluded patterns, separated by ;
CVI_MCP_ALLOW_BODY 0 1 allows function bodies out
CVI_MCP_ALLOW_PEEK 0 1 allows short excerpts with peek
CVI_MCP_MAX_SNIPPET_LINES 40 line cap per snippet
CVI_MCP_BUILD_CONFIG first Debug .prj configuration replayed by the preprocessor and build_check
CVI_MCP_DEFINES empty extra or forced macros: OPTION=1;OTHER
CVI_MCP_INDEX_UNLISTED 0 1 also indexes .c and .h of the folder missing from the .prj (in_build: false)
CVI_MCP_MAX_CHARS 40000 size budget of one answer
CVI_MCP_FORMAT json compact: lists as columns and rows
CVI_MCP_AUDIT_PATHS 1 0 hashes file names in the audit log
CVI_MCP_SETTINGS .cvi-mcp.json project settings file
CVI_MCP_SDK_PATHS auto NI SDK folders, separated by ;
CVI_MCP_BUILD_CMD auto path to the CVI build executable
CVI_MCP_BUILD_ARGS {project} argument template, {project} and {target}
CVI_MCP_BUILD_TIMEOUT 300 build timeout in seconds

Exclusions match both the path relative to the project root and the bare file name, so vendor/* and *_generated.c both work. Use them for a third-party tree you do not want indexed, or for files no agent should ever read.

An optional .cvi-mcp.json next to the .prj tunes analyses for one project. Every key is optional:

{
  "build_config": "Release64",
  "defines": {"NOMBRE_MAX": "12"},
  "domains": {"camera": ["*Camera*", "Pylon*"], "plc": ["*Wago*", "*Modbus*", "MBT*"]},
  "resource_pairs": [["PylonDeviceOpen", "PylonDeviceClose", "arg0"]],
  "checked_calls": ["MyDriverWrite"],
  "thresholds": {"lines": [100, 200, 400], "cyclomatic": [10, 20, 50]}
}

resource_pairs adds acquire and release functions to the leak tracking; the third field says where the handle is: return (assigned result) or argN.

How much code actually leaves

With the defaults, an agent sees that StartCallback exists in main.c at line 13, takes six parameters, is called by nobody, and calls CommsInit then SetCtrlVal. It does not see a single line of the body.

That is enough to navigate and to map an architecture. It is not enough to spot a bug inside a function. Set CVI_MCP_ALLOW_BODY=1 when you want real analysis, keep the line cap at a value you are comfortable with, and read ~/.cvi-mcp/audit.jsonl whenever you want to know exactly what went out: every call is logged with the bytes returned and the lines of code that left.

build_check is the one tool that runs something. It invokes the CVI build command and parses the output. It applies no fix and touches no source, but a compilation produces its usual output files.

English

Français

Clone this wiki locally