Use with AI Coding Agents¶
Every MCP client needs the same three things: a command to spawn, arguments,
and the DRUPAL_* environment variables. Assembling those by hand differs
per client, so the server ships a self-installer that does it for you.
One command¶
With no arguments it:
- Detects the MCP clients on this machine (Claude Code, Claude Desktop,
Cursor, Codex CLI, OpenCode, Windsurf, Pi, Hermes Agent, OpenClaw) — by
finding their CLI on
PATHor a known config path. - Gathers
DRUPAL_*settings from./.env, the process environment and any--envflags (precedence: dotenv < environment < flags). - Writes each client's config — preferring the client's own
mcp addcommand when it has one, otherwise writing the documented file directly.
✓ claude-code: Added stdio MCP server drupal ... to local config
✓ codex: Added global MCP server 'drupal'.
✓ opencode: MCP server "drupal" added to ./opencode.json
Useful options¶
| Option | Effect |
|---|---|
--client NAME |
Configure only this client (repeatable) |
--all |
Configure every supported client |
--list |
List supported client names |
--print |
Print the config/commands without writing anything |
--json |
Emit results as JSON (for scripted/agent-driven setup) |
--env KEY=VALUE |
Extra environment variable (repeatable) |
--dotenv PATH |
Env file to read (default ./.env if present) |
--name NAME |
Server name inside the client config (default drupal) |
--command CMD |
Launch command override, e.g. --command 'uvx --from drupal-mcp-server drupal-mcp' |
--scope local\|user\|project |
Claude Code configuration scope |
--global |
OpenCode: write the global config instead of the project one |
What gets written where¶
| Client | Config | Mechanism |
|---|---|---|
| Claude Code | claude mcp add (or ./.mcp.json with --scope project) |
client CLI |
| Claude Desktop | claude_desktop_config.json (mcpServers) |
direct write |
| Cursor | ~/.cursor/mcp.json (mcpServers) |
direct write |
| Codex CLI | codex mcp add (or ~/.codex/config.toml) |
client CLI |
| OpenCode | opencode mcp add (or opencode.json → mcp.servers) |
client CLI |
| Windsurf | ~/.codeium/windsurf/mcp_config.json (mcpServers) |
direct write |
| Pi | pi mcp add (or ~/.pi/agent/mcp.json) |
client CLI |
| Hermes Agent | ~/.hermes/config.yaml (mcp_servers) |
direct write |
| OpenClaw | openclaw mcp add |
client CLI (required) |
The launch command is the absolute path to the drupal-mcp script when it is
on PATH (GUI clients like Claude Desktop launch servers with a sanitized
environment, so a bare name can fail there); otherwise
uvx --from drupal-mcp-server drupal-mcp runs it straight from PyPI.
Existing configs are merged, never clobbered: other servers in the same file survive, a server of the same name is replaced (so re-running is idempotent), and a config that fails to parse is left untouched with an error instead of being overwritten.
Manual setup (copy-paste)¶
If you would rather wire it up yourself — every command below is verified against the client's current CLI or docs:
openclaw mcp add drupal \
--command drupal-mcp \
--env DRUPAL_BASE_URL=https://your-site.example \
--env DRUPAL_USERNAME=admin \
--env DRUPAL_PASSWORD=secret \
--no-probe
OpenClaw's mcp add refuses to run while the local Gateway is up — stop it
first, then verify with openclaw mcp doctor drupal --probe.
No local install?¶
Point any client at PyPI directly:
Verify¶
After installing, restart the client and ask it something that must touch
Drupal, for example "list my content types" (the agent should call
inventory_content_types). Client-side checks also help:
- Claude Code:
/mcpshows the server and its connection state - OpenClaw:
openclaw mcp status --verbose, thenopenclaw mcp probe drupal - Hermes:
hermes mcp test drupal - Pi:
pi mcp list
Security notes¶
The written config contains your DRUPAL_* values, including the password.
Project-scope files that live inside a repository (.mcp.json,
opencode.json) can end up committed — the installer warns when it writes
credentials into the current directory. Prefer --scope local for Claude
Code (stored outside the repo) or keep secrets in ~/.env-style files
outside the project and reference them from the client yourself.