LLM Client Configuration

This guide provides settings for connecting your AI client to Capella MCP. For installation, the user manual and release notes, see the Obeo AI for Capella documentation.

Install the product using the installation guide first. You need an agentic AI client that can add a custom MCP server and invoke its tools. The client combinations tried cover CLIs, desktop applications and IDE integrations. Check that your chosen client supports the local command or server URL used by your connection mode. Choose a transport and server profile before applying the client-specific examples below. These settings retain the client versions and limitations recorded in the original guides; client UI labels can change. The desktop quick start uses the default edit profile; examples explicitly requesting read-only restrict the available tools.

See Transport and access rights for process permissions, connection approvals and the account used for model operations.

Examples use Windows launchers by default. Replace <installation-folder-absolute-path> with your absolute installation path and x.y.z.YYYYMMDDHHMM (or <version>) with your extracted agent version. On Linux/macOS, use forward slashes and capella-mcp instead of capella-mcp.cmd; invoke the launcher directly rather than through cmd.exe. See Paths and executable names for the Java, Capella, CLI and installer equivalents. Escape Windows backslashes in JSON and double-quoted TOML strings.

1. Quick-start ChatGPT desktop settings

These labels match version 26.911.61220, released on September 16, 2026.

Open Settings > Plugins > MCP and select Add MCP server.
Enter capella-mcp as the name and select STDIO.
Enter the full launcher path for your platform below. Replace <installation-folder-absolute-path> with your absolute installation path and <version> with the version of the extracted agent folder.
Leave arguments empty. Save and restart the client.

Windows

<installation-folder-absolute-path>\tools\sirius-automation-agent-<version>\bin\capella-mcp.cmd

Linux or macOS

<installation-folder-absolute-path>/tools/sirius-automation-agent-<version>/bin/capella-mcp

For a server configured with STDIO, the desktop client starts the MCP server automatically.

For other clients or connection methods, see the installation guide.

2. Stdio Mode

2.1. ChatGPT Desktop App Stdio

The following menu labels match ChatGPT desktop app version 26.911.61220, released September 16, 2026:

Open Settings.
Select Plugins.
Select the MCP tab.
Select Add MCP server.
Enter capella-mcp as the name and choose STDIO as the type.
Enter the full Windows or Linux/macOS command path shown above.
Leave server arguments empty, or set --profile read-only for a restricted check.
Save and restart the client.

For a server configured with STDIO, the desktop client starts the MCP server automatically. Return to Settings > Plugins > MCPs to confirm that the server is enabled. See OpenAI’s MCP documentation for general configuration; menu labels can differ by app version: https://learn.chatgpt.com/docs/extend/mcp

2.2. Codex CLI Stdio

ChatGPT desktop app, Codex CLI, and the Codex IDE extension share MCP server configuration in ~/.codex/config.toml. See OpenAI’s current MCP documentation: https://learn.chatgpt.com/docs/extend/mcp

Add the local Capella MCP server with the Codex CLI. In these Windows examples, replace <installation-folder-absolute-path> with your absolute installation path and x.y.z.YYYYMMDDHHMM with your extracted agent version:

codex mcp add capella-mcp -- "<installation-folder-absolute-path>\tools\sirius-automation-agent-x.y.z.YYYYMMDDHHMM\bin\capella-mcp.cmd" --profile read-only
codex mcp list

Equivalent generic configuration:

[mcp_servers.capella-mcp]
command = "<installation-folder-absolute-path>\\tools\\sirius-automation-agent-x.y.z.YYYYMMDDHHMM\\bin\\capella-mcp.cmd"
args = ["--profile", "read-only"]
startup_timeout_sec = 20
tool_timeout_sec = 120

2.3. Antigravity IDE Stdio

Antigravity IDE can read MCP server definitions from ~/.gemini/config/mcp_config.json. See Google’s current MCP documentation: https://antigravity.google/docs/mcp In this Windows example, replace <installation-folder-absolute-path> with your absolute installation path and x.y.z.YYYYMMDDHHMM with your extracted agent version:

{
  "mcpServers": {
    "capella-mcp-stdio": {
      "command": "<installation-folder-absolute-path>\\tools\\sirius-automation-agent-x.y.z.YYYYMMDDHHMM\\bin\\capella-mcp.cmd",
      "args": ["--profile", "read-only"]
    }
  }
}

After restarting or reloading Antigravity IDE, open the agent side panel menu and select MCP Servers to inspect the server. Project permissions can be managed from Antigravity IDE’s project settings.

2.4. Claude Code Stdio

claude mcp add --transport stdio capella-mcp -- "<installation-folder-absolute-path>\tools\sirius-automation-agent-x.y.z.YYYYMMDDHHMM\bin\capella-mcp.cmd" --profile read-only

See Claude Code MCP configuration. Once connected, follow the guidance and connection check. You can also use the optional startup prompt to insert the initial Capella instructions.

2.5. Claude Desktop Stdio

Merge this entry into the existing mcpServers object in claude_desktop_config.json, preserving other servers, then restart Claude Desktop. Use the absolute launcher path for your installation:

{
  "mcpServers": {
    "capella-mcp": {
      "command": "<installation-folder-absolute-path>\\tools\\sirius-automation-agent-x.y.z.YYYYMMDDHHMM\\bin\\capella-mcp.cmd",
      "args": ["--profile", "read-only"]
    }
  }
}

See the official local-server setup for the configuration file location. Follow the guidance and connection check; if its fallback is needed, supply it in the chat. Do not assume Claude Desktop loads a Claude Code CLAUDE.md file.

2.6. VS Code on Windows (stdio)

This configuration was reported working with VS Code Chat on Windows on September 18, 2026. Starting capella-mcp.cmd directly from the server setup failed in that trial; launching it through cmd.exe with an explicit working directory resolved startup.

Open the Chat view, then Settings > MCP Servers > Add Server, and choose stdio. These labels reflect the trial; if they differ in your version, run MCP: Add Server from the Command Palette.
Open the server configuration: use .vscode/mcp.json for the workspace, or MCP: Open User Configuration for your user profile.
Merge the following entry into the top-level servers object, preserving other servers. Replace cwd with the absolute path to your extracted agent’s bin directory. Replace <installation-folder-absolute-path> with your absolute installation path and x.y.z.YYYYMMDDHHMM with your extracted agent version.

{
  "servers": {
    "capella-mcp-stdio": {
      "type": "stdio",
      "command": "cmd.exe",
      "cwd": "<installation-folder-absolute-path>\\tools\\sirius-automation-agent-x.y.z.YYYYMMDDHHMM\\bin",
      "args": [
        "/d",
        "/c",
        ".\\capella-mcp.cmd",
        "--transport",
        "stdio"
      ]
    }
  }
}

The cwd value places the process in the launcher’s directory, so .\capella-mcp.cmd resolves there. Keep Windows backslashes doubled inside JSON strings. This example uses the default edit profile; see Server Profiles to restrict the available tools.

Save the configuration and restart the server. Use MCP: List Servers, select capella-mcp-stdio, then Show Output to inspect startup errors and tool discovery. The recorded trial reported Discovered 43 tools; the count depends on the product version and profile and is not a required result. Tool discovery confirms the MCP connection, but does not verify access to the running Capella instance. Continue with the guidance and connection check.

Use stdio for this setup. The trial also reported an HTTP connection issue suspected to involve the legacy SSE transport; its cause was not confirmed. Capella MCP’s HTTP endpoint uses Streamable HTTP, not the legacy SSE transport. Current VS Code documentation describes HTTP Stream support with fallback to SSE, so this observation is not a general VS Code HTTP incompatibility. VS Code HTTP with Capella is not validated by this stdio trial.

See VS Code MCP setup and diagnostics for current configuration and logging commands.

3. Local HTTP Mode

Start the server as described in the installation guide before configuring the client.

3.1. ChatGPT Desktop App HTTP

First start capella-mcp in local HTTP mode as described above. The following menu labels match ChatGPT desktop app version 26.911.61220, released September 16, 2026:

Open Settings.
Select Plugins.
Select the MCP tab.
Select Add MCP server.
Enter capella-mcp as the name and choose Streamable HTTP as the type.
In URL, enter http://127.0.0.1:3333/mcp.
Save and restart the client.

Return to Settings > Plugins > MCPs to confirm that the server is enabled. See OpenAI’s MCP documentation for general configuration; menu labels can differ by app version: https://learn.chatgpt.com/docs/extend/mcp

4. HTTP bearer token

If the server uses CAPELLA_MCP_AUTH_TOKEN (or AUTOMATION_MCP_AUTH_TOKEN), configure the client to send the same bearer token on every request, including resource requests. For Codex HTTP, add bearer_token_env_var = "CAPELLA_MCP_AUTH_TOKEN" to the server’s TOML table and make that variable available to the client process too.

5. MCP Client Examples

The examples below use local HTTP. For stdio examples, see Stdio Mode above.

5.1. Codex CLI HTTP

Add the local Capella MCP server with the Codex CLI:

codex mcp add capella-mcp --url http://127.0.0.1:3333/mcp
codex mcp list

Example HTTP configuration:

[mcp_servers.capella-mcp]
url = "http://127.0.0.1:3333/mcp"
startup_timeout_sec = 20
tool_timeout_sec = 120

5.2. Claude Code HTTP

Claude Code supports MCP server configuration through its MCP setup flow and project configuration. See the Claude Code MCP docs: https://code.claude.com/docs/en/mcp

Add the local Capella MCP server with Claude Code:

claude mcp add --transport http capella-mcp http://127.0.0.1:3333/mcp
claude mcp list

Once connected, you can use the optional startup prompt to insert the initial Capella instructions. Follow the guidance and connection check before modeling.

Example project .mcp.json configuration:

{
  "mcpServers": {
    "capella-mcp": {
      "type": "http",
      "url": "http://127.0.0.1:3333/mcp"
    }
  }
}

The local HTTP endpoint is:

http://127.0.0.1:3333/mcp

5.3. Antigravity CLI HTTP

Antigravity CLI reads global MCP server definitions from ~/.gemini/config/mcp_config.json and workspace definitions from .agents/mcp_config.json. See Google’s current MCP documentation: https://antigravity.google/docs/mcp

Example global configuration:

{
  "mcpServers": {
    "capella-mcp": {
      "serverUrl": "http://127.0.0.1:3333/mcp"
    }
  }
}

Start Antigravity CLI from the modeling work folder:

agy

Then enter /mcp in the Antigravity CLI prompt to verify that capella-mcp is connected. Antigravity CLI requires serverUrl for Streamable HTTP and SSE servers; the legacy url and httpUrl keys are not supported.

5.4. Gemini CLI HTTP - Enterprise/API

Google transitioned unpaid and Google One terminal users from Gemini CLI to Antigravity CLI. Gemini CLI remains available for supported enterprise, Google Cloud, and paid API-key deployments; use this section only for those contexts.

Gemini CLI can use MCP servers from its settings. See: https://geminicli.com/docs/tools/mcp-server/

Add the local Capella MCP server with the Gemini CLI:

gemini mcp add --transport http capella-mcp http://127.0.0.1:3333/mcp
gemini mcp list

Example settings.json configuration:

{
  "mcpServers": {
    "capella-mcp": {
      "httpUrl": "http://127.0.0.1:3333/mcp",
      "timeout": 30000
    }
  }
}

Use httpUrl for streamable HTTP. Gemini CLI uses url for SSE servers.

The local HTTP endpoint is:

http://127.0.0.1:3333/mcp

5.5. OpenCode HTTP

OpenCode defines MCP servers under mcp in its configuration. See: https://opencode.ai/docs/mcp-servers/

Example opencode.json configuration:

{
  "$schema": "https://opencode.ai/config.json",
  "mcp": {
    "capella-mcp": {
      "type": "remote",
      "url": "http://127.0.0.1:3333/mcp",
      "enabled": true
    }
  }
}

The local HTTP endpoint is:

http://127.0.0.1:3333/mcp