MCP Tools

The MCP server’s tools are the CranL operations, with the same names, inputs and errors as the REST API, CLI and SDKs. For example deploy_app, get_app_logs and create_sandbox. Every operation is listed in the Operation Reference.

Choosing tools

URL	Tools loaded
`/api/mcp`	Default toolsets: `core`, `apps`, `deployments`, `observability`, `env`, `domains`, `databases`, `sandboxes`, plus `call_operation`
`/api/mcp?toolsets=all`	Every operation
`/api/mcp?toolsets=email,storage`	`core` plus the listed toolsets
`?readonly=1`	Read-only operations only
`/api/mcp?toolsets=sandboxes&readonly=1`	The read-only operations of `core` plus `sandboxes`

readonly combines with toolsets. Toolset names are listed in the Operation Reference.

Reaching tools that aren’t loaded

You don’t need to load everything. Two tools reach the rest:

Tool	What it does
`find_tools`	Searches every CranL operation by what you want to do (“run a migration”, “redirect www”, “invite a teammate”, “follow logs”) and returns the best matches with their input schemas.
`call_operation`	Runs any operation by name.

So an assistant connected with the defaults can still send email: it calls find_tools to find send_email and its input schema, then call_operation to run it.

call_operation takes the operation name and its arguments:

{ "operation": "run_command", "arguments": { "app_id": "…", "command": "npm run migrate" } }

operation is required.
arguments holds every field of that operation, including confirm, plan and dry_run.
Use find_tools first to get the operation’s input schema.

Scopes and safety

Each tool needs the scope listed in the reference. With an API key, only the tools its scopes allow will succeed. Others fail with insufficient_scope.
Destructive tools (⚠️ in the reference) return confirmation_required first. The assistant repeats the call with the confirm_token, or with request_approval: true to ask the owner. See Safety Controls.
Errors include a hint and next_actions, so the assistant knows what to do next. See Errors.

Other features

Progress: long calls such as deploys and sandbox creation report progress while they run.
Images: screenshots (sandbox_screenshot, sandbox_browser) and charts come back to the model as images.
Docs: the docs are available as cranl://docs/{topic} resources. The search_docs tool searches them.

Old tool names

The 17 old cranl_* tools still work. They’re hidden aliases for the new operations, so existing setups don’t break. Use the new operation names in new prompts, scripts and configs.

Old tool	New operation
`cranl_list_projects`	`list_projects`
`cranl_list_apps`	`list_apps`
`cranl_list_github_repos`	`list_github_repos`
`cranl_create_app`	`create_app` (`repository_id` is now `repository`)
`cranl_deploy_app`	`deploy_app`
`cranl_app_lifecycle`	`app_lifecycle`
`cranl_get_app_logs`	`get_app_logs`
`cranl_get_deployment_logs`	`get_deployment_logs`
`cranl_get_env`	`get_env`
`cranl_set_env`	`set_env`
`cranl_create_database`	`create_database`
`cranl_list_databases`	`list_databases`
`cranl_list_regions`	`list_regions`
`cranl_get_deployments`	`list_deployments`
`cranl_list_domains`	`list_domains`
`cranl_purge_cache`	`purge_cache`
`cranl_get_ai_fix`	`get_ai_fix`