strapi-mcp-capabilities

Strapi MCP Capabilities

Register custom MCP tools/prompts/resources so AI clients can trigger plugin-specific actions on a running Strapi instance. Strapi 5.47.0+. API source of truth: @strapi/types, search mcp.

Hard rules

1. Enable the server: config/server.ts must set mcp: { enabled: true }. Endpoint is POST /mcp.
2. Register only in the register() phase. The capability list is locked once the MCP server starts (during bootstrap). Registering later throws.
3. Use Zod from @strapi/utils (import { z } from '@strapi/utils'), not standalone zod.
4. Never use global strapi at module top-level. Either register inline (have strapi) or use ai.mcp.defineTool(...) to define at module scope, then register with the injected strapi.

Two registration patterns

Inline — when strapi is in scope (a function called from register()):

// mcp/tools/get-store-kpis.ts
import type { Core } from '@strapi/strapi';
import { z } from '@strapi/utils';

export const registerGetStoreKpisTool = (strapi: Core.Strapi) => {
  strapi.ai.mcp.registerTool({ /* definition */ });
};

Builder — define at module scope (no strapi), register later. defineTool is an identity helper that narrows the access variant for type inference:

// mcp/tools/get-top-products.ts
import { ai } from '@strapi/strapi';
import { z } from '@strapi/utils';

export const getTopProductsTool = ai.mcp.defineTool({ /* definition */ });

// register.ts
strapi.ai.mcp.registerTool(getTopProductsTool);

Wire both from the plugin register():

// register.ts
export const register = ({ strapi }: { strapi: Core.Strapi }) => {
  registerGetStoreKpisTool(strapi);          // inline
  strapi.ai.mcp.registerTool(getTopProductsTool); // builder
};

Tool definition (the core unit)

strapi.ai.mcp.registerTool({
  name: 'get_store_kpis',          // snake_case, unique
  title: 'Get Store KPIs',
  description: 'What it returns + when to use it. The AI reads this to choose the tool.',
  auth: { policies: [{ action: 'plugin::store-analytics.read' }] }, // OR devModeOnly: true
  resolveInputSchema: () => inputSchema,   // optional; omit if no args
  resolveOutputSchema: () => outputSchema, // required; a z.object
  createHandler: (strapi, context) => async ({ args, extra }) => {
    const result = await strapi
      .plugin('store-analytics')
      .service('analytics')
      .getStoreKpis(args);

    return {
      content: [{ type: 'text', text: JSON.stringify(result, null, 2) }],
      structuredContent: result, // must match resolveOutputSchema
    };
  },
});

Key points:

• createHandler(strapi, context) => handler. The handler receives a single object { args, extra }. args is z.infer<inputSchema> (typed); if no resolveInputSchema, omit args.
• context: { userAbility, user } — use userAbility for entity/field permission checks and user (token owner) for setCreatorFields on writes.
• Keep handlers thin: delegate to a plugin service; the tool is just the MCP surface.
• name/title/description/resolveOutputSchema are required. description quality drives AI tool selection — state what it returns and when to use it.

Schemas

• Define z.object input/output schemas; reuse shared field schemas (e.g. an isoDateSchema with .describe(...)).
• .describe() every non-obvious field — descriptions reach the AI client.
• structuredContent returned by the handler must satisfy resolveOutputSchema.

Auth (choose exactly one access variant)

• auth: { policies: [{ action, subject? }] } — CASL check; session passes if the token's ability satisfies any policy. action is typically a registered admin permission, e.g. plugin::<plugin-id>.<permission>.
• devModeOnly: true — no auth; only available when autoReload is on (dev). Use for debug tools.

For auth to be grantable to a token, register the permission action in bootstrap():

await strapi.service('admin::permission').actionProvider.registerMany([
  { section: 'plugins', displayName: 'Access store analytics',
    uid: 'read', pluginName: 'store-analytics' },
]);
// → action becomes 'plugin::store-analytics.read'

Handler return shape

// success
return { content: [{ type: 'text', text }], structuredContent: result };
// error
return { content: [{ type: 'text', text: 'message' }], isError: true };

structuredContent and isError are mutually exclusive.

Folder convention

server/src/
├── register.ts            # calls registrations (register phase only)
├── bootstrap.ts           # registerMany permission actions
├── mcp/
│   ├── index.ts           # registerAllMcpTools(strapi)
│   ├── schemas.ts         # shared zod field schemas
│   └── tools/<tool>.ts    # one capability per file
└── services/              # business logic the handlers call

Verify

1. mcp: { enabled: true } in config/server.ts; restart Strapi.
2. Connect a client to http://localhost:1337/mcp with Authorization: Bearer <admin-token> whose permissions include the tool's auth action.
3. Confirm the tool appears and returns structuredContent matching its output schema.

Prompts & resources

Same lifecycle/auth rules, different builders: ai.mcp.definePrompt / strapi.ai.mcp.registerPrompt and ai.mcp.defineResource / strapi.ai.mcp.registerResource. See reference.md for their field tables and signatures.

Docs

• MCP server (config, permission boundaries, stateless architecture, limitations)
• API types: @strapi/types