Connecting any MCP client
There is nothing Claude-specific about Veri's server: it's a standard MCP server speaking stdio. Any client that can launch a stdio server — an agent CLI, an IDE, your own tooling — connects with the same three facts. Using one of the agents Veri configures automatically? See Claude Code, Cursor, Codex CLI, or Gemini CLI instead — Veri writes those configs for you.
The three facts
| Fact | Value |
|---|---|
| Command | node — your own Node, version 20 or newer |
| First argument | The server script that ships inside the Veri app:/Applications/Veri.app/Contents/Resources/app/node_modules/@verikb/mcp/dist/server.js(no version segment — the path survives app self-updates). From a source checkout it's packages/mcp/dist/server.js after npm install && npm test. |
| Second argument | The project root the server serves: the directory containing veri/. Relative paths resolve against it. |
One server serves one project. The server is subprocess-free, reads and writes only the project's files, and makes no network calls.
A generic config
Most clients accept the common mcpServers JSON shape — this is exactly what Veri's own connection panel writes as a project-scoped .mcp.json:
{
"mcpServers": {
"veri": {
"command": "node",
"args": [
"/Applications/Veri.app/Contents/Resources/app/node_modules/@verikb/mcp/dist/server.js",
"/Users/you/code/your-project"
]
}
}
}Clients with their own format (TOML, YAML, a settings UI) need the same three facts in their own spelling: a stdio server named veri, command node, and the two arguments above. If your client distinguishes transports, choose stdio — Veri's server has no HTTP or SSE mode.
Verify it works
The most direct proof doesn't need the client at all: Veri's connection panel (Settings → Agent connection → Verify connection) launches the server with the exact command and arguments and performs a real MCP handshake, including a search call proving it serves this project's documents. If that's green, any remaining problem is in your client's config file, not the server. Failures name one of five causes, each covered in troubleshooting.
What your client gets
The same eleven tools every connected agent gets — get_context, get_import_instructions, search, get_document, get_neighbors, file_decision, file_work_order, file_requirement, file_source, file_receipt, run_check — detailed in the tool table on the Claude Code page. Everything a tool writes lands as a proposed/draft markdown file for your review; nothing binds without your approval stamp.