Connect AI coding tools

Choose how

Skills teach, MCP acts. Most people want both, which is what the plugin is. SDODS is free for all of them: Apache-2.0 on npm, unlimited API tokens, no paid tier.

Claude Code plugin

/plugin marketplace add siri1410/sdods-skills

/plugin install sdods@sdods

To offer the plugin to everyone who opens a repository, commit this to .claude/settings.json. Claude Code asks each person to trust the marketplace the first time.

{
  "extraKnownMarketplaces": {
    "sdods": { "source": { "source": "github", "repo": "siri1410/sdods-skills" } }
  },
  "enabledPlugins": { "sdods@sdods": true }
}

Update with /plugin marketplace update sdods.

Skills

The skills follow the open Agent Skills format, so one copy works across agents. Install them with the skills CLI, which asks which agents you use:

npx skills add siri1410/sdods-skills

Add -a claude-code, -a codex, -a cursor or -a github-copilot to pick agents up front, and -g to install for your user instead of the project.

Or from npm, without GitHub:

npx -y @sdods/cli skills install

That writes .claude/skills and .agents/skills; --agent cursor,copilot,gemini adds other folders and --global installs under your home directory. See sdods skills.

Local MCP server

Every stdio client runs the same command. There is no npm package called sdods; the command belongs to @sdods/cli, and -y lets npx fetch it without a prompt:

npx -y @sdods/cli mcp --project <slug> --env <env>

--project and --env are defaults for tools and prompts, and both are optional. The server finds the workspace from the directory the client starts it in. A client that starts servers somewhere else — Claude Desktop, or any user-level config — needs the workspace named:

npx -y @sdods/cli --cwd /path/to/workspace mcp

For the clients SDODS knows, one command writes the configuration and keeps any servers already there:

sdods mcp install claude

The other targets are codex, cursor, vscode, windsurf and gemini.

Dedicated MCP endpoint

Every SDODS server serves MCP over Streamable HTTP at /mcp: sdods serve on a machine your team can reach, the Docker image or a Cloud Run deployment. Nothing runs on each laptop, runs and results are shared, and each person's token decides which tools they see.

sdods tokens create --user <username> --name claude-code --scopes runs:read,runs:write,features:read

export SDODS_TOKEN=amx_...

curl https://<your-sdods-host>/api/mcp/info

If your administrator gave you an account on https://api.sdods.com, the endpoint is https://api.sdods.com/mcp. The snippets below use https://<your-sdods-host>/mcp.

Every client

Each section has the local (stdio) form first and the endpoint form second.

Claude Code

claude mcp add sdods -- npx -y @sdods/cli mcp

claude mcp add --transport http sdods https://<your-sdods-host>/mcp --header 'Authorization: Bearer ${SDODS_TOKEN}'

Add -s project to write .mcp.json for the whole team. The single quotes keep ${SDODS_TOKEN} in the file, and Claude Code expands it from each person's environment. Check with /mcp.

Codex

codex mcp add sdods -- npx -y @sdods/cli mcp

codex mcp add sdods --url https://<your-sdods-host>/mcp --bearer-token-env-var SDODS_TOKEN

Both write ~/.codex/config.toml, which the Codex CLI, IDE extension and desktop app share:

[mcp_servers.sdods]
url = "https://<your-sdods-host>/mcp"
bearer_token_env_var = "SDODS_TOKEN"

Gemini CLI

gemini mcp add sdods npx -- -y @sdods/cli mcp

gemini mcp add --transport http --header "Authorization: Bearer $SDODS_TOKEN" sdods https://<your-sdods-host>/mcp

The -- matters: without it Gemini reads -y as its own flag. In settings.json, httpUrl is Streamable HTTP; url would mean the older SSE transport.

{
  "mcpServers": {
    "sdods": {
      "httpUrl": "https://<your-sdods-host>/mcp",
      "headers": { "Authorization": "Bearer $SDODS_TOKEN" }
    }
  }
}

Cursor

Use Add to Cursor in the box at the top of this page, or write the file:

{
  "mcpServers": {
    "sdods": { "command": "npx", "args": ["-y", "@sdods/cli", "mcp"] }
  }
}

{
  "mcpServers": {
    "sdods": {
      "url": "https://<your-sdods-host>/mcp",
      "headers": { "Authorization": "Bearer ${env:SDODS_TOKEN}" }
    }
  }
}

~/.cursor/mcp.json makes it available in every project.

VS Code and GitHub Copilot

Use Add to VS Code at the top of this page, run code --add-mcp '{"name":"sdods","command":"npx","args":["-y","@sdods/cli","mcp"]}', or write the file. VS Code calls the top-level key servers, and prompts for the token once with an input:

{
  "inputs": [
    {
      "id": "sdods-token",
      "type": "promptString",
      "description": "SDODS API token",
      "password": true
    }
  ],
  "servers": {
    "sdods": {
      "type": "http",
      "url": "https://<your-sdods-host>/mcp",
      "headers": { "Authorization": "Bearer ${input:sdods-token}" }
    }
  }
}

For stdio, the server entry is { "type": "stdio", "command": "npx", "args": ["-y", "@sdods/cli", "mcp"] }.

Windsurf

Windsurf reads one user-level file and names the remote URL serverUrl:

{
  "mcpServers": {
    "sdods": {
      "serverUrl": "https://<your-sdods-host>/mcp",
      "headers": { "Authorization": "Bearer ${env:SDODS_TOKEN}" }
    }
  }
}

For stdio use { "command": "npx", "args": ["-y", "@sdods/cli", "--cwd", "/path/to/workspace", "mcp"] }. Windsurf caps the number of enabled tools; --caps core,run,analyze keeps SDODS under it.

Zed

{
  "context_servers": {
    "sdods": { "command": "npx", "args": ["-y", "@sdods/cli", "mcp"] }
  }
}

For the endpoint: { "url": "https://<your-sdods-host>/mcp", "headers": { "Authorization": "Bearer <token>" } }.

JetBrains AI Assistant and Junie

In AI Assistant, open Settings → Tools → AI Assistant → Model Context Protocol, add a server and paste the mcpServers JSON from the Cursor section. Junie reads the same shape from .junie/mcp/mcp.json in the project or ~/.junie/mcp/mcp.json.

OpenCode

{
  "mcp": {
    "sdods": { "type": "local", "command": ["npx", "-y", "@sdods/cli", "mcp"], "enabled": true }
  }
}

For the endpoint: { "type": "remote", "url": "https://<your-sdods-host>/mcp", "headers": { "Authorization": "Bearer <token>" }, "oauth": false }.

Amp

amp mcp add sdods -- npx -y @sdods/cli mcp

Kiro

{
  "mcpServers": {
    "sdods": { "command": "npx", "args": ["-y", "@sdods/cli", "mcp"] }
  }
}

Cline, Roo Code, Continue, Goose, Warp and Factory Droid

Every one of them takes npx -y @sdods/cli mcp for stdio.

Claude Desktop

Claude Desktop's config file starts stdio servers only, outside any project, so name the workspace:

{
  "mcpServers": {
    "sdods": {
      "command": "npx",
      "args": ["-y", "@sdods/cli", "--cwd", "/path/to/workspace", "mcp"]
    }
  }
}

The file is ~/Library/Application Support/Claude/claude_desktop_config.json on macOS and %APPDATA%\Claude\claude_desktop_config.json on Windows. To reach an endpoint instead, bridge it with mcp-remote:

{
  "mcpServers": {
    "sdods": {
      "command": "npx",
      "args": [
        "-y",
        "mcp-remote",
        "https://<your-sdods-host>/mcp",
        "--header",
        "Authorization:${AUTH_HEADER}"
      ],
      "env": { "AUTH_HEADER": "Bearer <token>" }
    }
  }
}

LM Studio uses the same mcpServers shape in ~/.lmstudio/mcp.json.

Check it works

List the tools without any client involved — this is also the fastest way to see a real error:

npx -y @sdods/cli mcp --list-tools

Then ask the agent something only SDODS can answer, such as "list the SDODS projects and their environments".

Next steps

• MCP — topology, tool families and scopes
• Claude Code and Codex — running SDODS agents through them
• MCP tools — the full catalogue