I added skills to PolyMCP to handle a common problem with MCP servers: once the number of tools grows, feeding raw schemas to agents becomes messy and inefficient.
Challenges we faced: • Agents consume too much context/tokens loading raw schemas. • Tool discovery becomes noisy and hard to manage. • Different agents need different subsets of tools. • Orchestration logic leaks into prompts.
PolyMCP introduces skills — curated, structured sets of tools with documentation, organized so agents can consume only what’s relevant.
PolyMCP handles tool organization, letting agents with Ollama,OpenAI and more to focus on execution.
Example: generate skills from a Playwright MCP server in one command:
polymcp skills generate --servers "npx @playwright/mcp@latest"
Benefits of skills: • Reuse capabilities across multiple agents. • Keep agent context small while scaling the number of tools. • Control what each agent can access without modifying the agent itself.
Repo: https://github.com/poly-mcp/Polymcp
Would love feedback on approaches for tool organization and minimizing context/token usage in multi-agent setups.