RationalBloks

RationalBloks MCP Server

Deploy production APIs in minutes. Tools for projects, schemas, deployments, modules, object storage, and graph data — delivered on infrastructure you own (self-host or your own BYOC cluster).

What Is This?

RationalBloks MCP lets AI agents (Claude Code, Claude Desktop, Cursor, VS Code, etc.) deploy production APIs from a JSON schema, and deploy and run the frontends and backends built on them. No backend code to write. No infrastructure to manage.

"Create a task management API with tasks, projects, and users"

→ 2 minutes later: Production API running on Kubernetes

Installation

# recommended — no install step; @latest checks for a new release at every start

uvx rationalbloks-mcp@latest

or install into an environment

pip install rationalbloks-mcp

Quick Start

1. Get Your API Key

Visit rationalbloks.com/settings and create an API key. A key carries scopes: a read key reaches only the read tools, a read + write key reaches every tool. The server lists only the tools your key may call.

2. Configure Your AI Agent

Claude Code — remote server (no install):

claude mcp add --transport http rationalbloks https://mcp.rationalbloks.com/mcp \

--header "Authorization: Bearer rb_sk_your_key_here"

or the local server:

claude mcp add --env RATIONALBLOKS_API_KEY=rb_sk_your_key_here rationalbloks \

-- uvx rationalbloks-mcp@latest

--scope project writes the server to the repository's .mcp.json instead, to share it with a team. Keep the key out of the file with "Authorization": "Bearer ${RATIONALBLOKS_API_KEY}": Claude Code expands the variable from each developer's environment.

VS Code — add to .vscode/mcp.json (Cursor: .cursor/mcp.json with a mcpServers key):

{

"servers": { "rationalbloks": { "command": "uvx", "args": ["rationalbloks-mcp@latest"], "env": { "RATIONALBLOKS_API_KEY": "rb_sk_your_key_here" } } } }

Claude Desktop — add to claude_desktop_config.json:

{

"mcpServers": { "rationalbloks": { "command": "uvx", "args": ["rationalbloks-mcp@latest"], "env": { "RATIONALBLOKS_API_KEY": "rb_sk_your_key_here" } } } }

Any other client: point it at https://mcp.rationalbloks.com/mcp (Streamable HTTP) with Authorization: Bearer rb_sk_....

3. Claude Code Permissions

Every tool that can lose data carries destructiveHint: true, and no other tool can lose data. One allow rule covers the server, and ask rules keep a prompt on the tools that can lose data. Claude Code evaluates ask rules before allow rules, so they prompt even though the server is allowed, and in auto mode they still prompt, because they run before the classifier. Add this to .claude/settings.json (the project) or ~/.claude/settings.json (every project):

{

"permissions": { "allow": ["mcprationalbloks"], "ask": [ "mcprationalbloks__update_schema", "mcprationalbloksdrop_schema_items", "mcprationalbloksdeploy_destructive", "mcprationalbloksrollback_project", "mcprationalbloksdelete_project", "mcprationalbloksupdate_graph_schema", "mcprationalbloksrollback_graph_project", "mcprationalbloksdelete_graph_project", "mcprationalbloksdelete_graph_node", "mcprationalbloksdelete_graph_relationship", "mcprationalbloksset_module_env", "mcprationalbloksdelete_module" ] } }

Add "mcp__rationalbloks__deploy_production" to ask to approve each production release yourself.

Asked before every change: allow only the read tools instead. These rules match every read tool and no other:

{

"permissions": { "allow": [ "mcprationalbloksget_*", "mcprationalblokslist_*", "mcprationalblokssearch_graph_nodes", "mcprationalbloksfulltext_search_graph", "mcprationalblokstraverse_graph", "mcprationalblokspreview_schema_change" ] } }

Read-only server: register a second server with a read key. It lists only the read tools and the platform refuses every other call, so one rule allows all of it:

claude mcp add --transport http rationalbloks-read https://mcp.rationalbloks.com/mcp \

--header "Authorization: Bearer rb_sk_your_read_key"

{ "permissions": { "allow": ["mcp__rationalbloks-read"] } }

Auto mode reads autoMode only from ~/.claude/settings.json, managed settings or --settings, never from a project's .claude/settings.json. If its classifier blocks routine RationalBloks calls, describe your project there, keeping $defaults (replace MyProject and apps.yourdomain.com with your project's name and your cluster's domain):

{

"autoMode": { "environment": [ "$defaults", "Key internal services: RationalBloks (MCP server rationalbloks, https://mcp.rationalbloks.com), where our project MyProject runs; its staging environment is for development", "Trusted internal domains: rationalbloks.com and *.rationalbloks.com; *.apps.yourdomain.com, where MyProject's APIs and modules are served", "Sensitive remote targets: MyProject's production environment" ] } }

Run claude auto-mode config to see the rules in effect, or edit them in /permissions (Auto mode tab).

---

Tools

Read tools only read. Write tools change something, and never lose data. Destructive tools can lose data, and carry destructiveHint: true: they replace a whole schema (update_schema, update_graph_schema), drop, delete, roll back, deploy a plan that drops data (deploy_destructive), or overwrite a module's environment (set_module_env).

A plan drops data when it drops a table or field, or changes a field's type in a way that rounds or cuts its stored values (a decimal given a smaller scale or turned integer, a datetime turned date). The deploy tools refuse such a plan and name each drop; deploy_destructive applies it.

A long operation (create, deploy, rollback, delete, module operations) runs as a job: the tool answers a job_id, and get_job_status follows it to its result.

Relational Projects

Graph Projects

Graph Data

Modules

A module is one of a project's own frontends (frontblok) or backends (logicblok), built from a GitHub repository and run beside the project.

|