AI
MCP and Skills: Two Ways to Give Agents Context
MCP is an integration protocol with three primitives: tools for actions, resources for context, and prompts for templates. Skills are folders containing instructions, scripts, and references. Unlike MCP tools with typed arguments and isolated processes, skills have no schema and require shell access. The key difference is purpose. MCP solves integration, connecting any agent to any service. Skills solve context saturation through progressive disclosure, letting the agent load only what it needs.
Typed Tools vs Flexible Scripts
MCP tools use JSON schemas for typed arguments, run in isolated processes, and can be called remotely. Their downside is schema bloat. Scripts in skills have no schema, iterate instantly, and agents can self-author them. But scripts need shell access and have portability issues across operating systems. MCP resources are app-driven and can load dynamic content remotely. Skills reference files are agent-friendly, work offline, and can be authored by the agent. Prompts are user-invoked templates; skills.md are agent-pulled with rich content and conditional logic.
Integration vs Context Saturation
MCP solves the integration problem: connect N agents with M services using one protocol. Skills solve the context problem: don’t load everything at once. Progressive disclosure lets the agent discover information as needed. Rodrigues used a racing metaphor: MCP is the pilot, skills are the navigator. You don’t fire the pilot because the co-pilot knows the way. The debate between skills and MCP is false. They complement each other.
Supabase Benchmark: MCP Plus Skills Wins
Supabase exposes an MCP server with over 20 tools. They designed a skill that incentivizes agents to search docs, includes a security checklist, and opinionated workflows for schema management. In a test, an agent with MCP only ignored the search docs tool and created a Postgres view that bypassed row-level security. Adding the skill made the agent use security invoker, keeping RLS intact. In a three-project benchmark, MCP plus skill outperformed all other conditions, including MCP-only and no-MCP baselines.
Q&A
How do skills differ from MCP resources? MCP resources are user-driven; skills are agent-pulled, allowing the agent to decide when to load context. ▶ Watch (20:25)
What powers the search docs tool in the MCP server? It’s built on a GraphQL API for documentation search, more sophisticated than web search. ▶ Watch (21:24)
Notable Quotes
skills are not the MCP killer Pedro Rodrigues · ▶ Watch (2:09)
MCP alone is not enough Pedro Rodrigues · ▶ Watch (16:57)
the bottleneck here is context, not capability tools Pedro Rodrigues · ▶ Watch (17:41)
Key Takeaways
- Skills complement MCP by solving context saturation through progressive disclosure.
- Agents default to training data and need guidance to search up-to-date documentation.
- Combining MCP and Skills outperforms MCP alone in security and task completion.
About the Speaker
Pedro Rodrigues is an AI Tooling Engineer at Supabase, maintaining their MCP server, AI assistant, and Skills. He has contributed to the MCP protocol since its early days, including work on its SDKs and the Skybridge project. He also co-founded the Lisbon AI Week conference.