MCP Resource

An MCP resource is the read only data slice that an MCP server publishes for AI models to consume. The server advertises its resources through a standard list method. The client then issues a read request with a specific URI and receives structured data the model can reason over. Resources stay completely passive. The model gets to see but never touch. This design choice matters because it lets teams expose sensitive information with confidence. In practice resources cover everything from Figma layer trees with exact positioning and variable bindings to Notion blocks with rich text and database relations to Postgres table schemas with column types and sample queries. The JSON-RPC nature means you can debug a resource read with a simple curl command against your local server. By 2026 this had become the standard way every serious AI client pulled context. Claude Code treated resources as first class. Cursor could read from six different servers in a single turn without breaking a sweat. The protocol won because resources gave models the right level of structure without forcing every tool maker to adopt one rigid schema.

An MCP resource is not a tool and it is not a prompt. Tools let the model take action like creating a new Figma component or posting to Slack. Prompts are reusable templates for common tasks. Resources only deliver data. It is not a screenshot either. Stop feeding Claude images of your Figma files. The model hallucinates measurements and misses nested variants. An MCP resource delivers the source truth with numbers the model can trust. It is not an unrestricted database query interface. Smart servers limit resources to specific views with role based filters so a design resource cannot accidentally expose your Stripe customer list. It is not the old plugin model from 2023 where every integration broke when the host updated. A resource server can evolve its internal logic while keeping the external contract stable. The distinction between resource and tool is what made MCP superior to every function calling spec that came before it.

A concrete example played out at a design tool company in spring 2026. They had spent months building a new illustration system but the handoff to engineering kept failing because specs lived in six places. The design lead stood up an MCP server for their Figma library an internal Notion knowledge base and their GitHub design tokens package. In one Claude Code session the engineer pasted a prompt asking the model to implement the new empty state illustration with correct tokens. The model listed resources across the three servers. It read the Figma frame resource and received the full component spec including stroke weights color stops text hierarchies and attached prototypes. It read the tokens resource from the GitHub server and pulled the exact semantic color mappings updated two days earlier. It read the Notion resource for the illustration guidelines page that explained the new tone of voice and three prior usability test findings. Then it generated the SVG React component and accompanying spec update in one shot. The engineer approved it without changes. The same task had taken two weeks of back and forth the previous quarter with endless screenshot threads and misread measurements. That single MCP workspace turned the design system from marketing slides into living code the model could navigate. Similar setups at Linear let their support agents read issue resources with full comment threads and at Stripe let ops agents read subscription resources with real conversion graphs pulled straight from production tables.

Use an MCP resource whenever you need the AI to work from truth instead of approximation. Designers should expose their Figma files as resources the minute they want the model to stop guessing sizes corner radii and variant states. Connect your design critique Notion database so the model can reference past feedback when suggesting changes to a checkout flow. Wire up your color typography and spacing tokens from their Git repo so every generated UI element matches your system automatically instead of requiring manual correction passes. Frontend developers used MCP resources from their component libraries to let the AI suggest refactors that stayed on brand with the model reading actual component props variants and usage examples before it touched the code. The pattern scales. One team connected eleven resources and started treating Claude like a senior designer who had read every file in the company. It works for audits too. Ask the model to verify a new screen against all documented accessibility rules after it reads the resources holding those rules. The 2026 teams that embraced resources early left the screenshot era behind and never looked back. They shipped tighter loops because the context was finally reliable.

Do not use an MCP resource when the task requires changing data. Read the resource then call a tool if you need to act. Skip resources for pure ideation sessions where the model should invent rather than analyze existing artifacts. They add no value when your entire context already fits in the base prompt. The biggest trap is weak auth. One high profile incident at a SaaS company in late 2025 occurred because their MCP resource for analytics did not respect tenant boundaries and the model mixed up customer data in a generated report. Always define resource scopes before you ship. Latency adds up too. Three sequential resource reads across different servers can push response time past ten seconds on slow connections. Prototype on your laptop with stdio but test the real HTTP plus SSE setup before you declare victory. Mobile remains a blind spot in 2026 so any workflow targeting phones should wait for better transport options.

MCP resources turn your fragmented design artifacts into one coherent context layer that AI can finally trust.