interpres: talk to any MCP server

AssemblyAI prices its Voice Agent API at $4.50 an hour and lists OpenAI Realtime at $18. Realtime can call MCP servers by itself. The Voice Agent API takes function and http tools only, so a team that switches for the price loses every MCP server its agent used to call. interpres gives them back. Type a website or paste an MCP URL. interpres looks in the official MCP registry, finds the server and turns each of its tools into a Voice Agent function tool. No website in mind? Search more than 11,000 public servers on the page and connect with one click. The agent sees ten tools at a time, since 45.6% of public servers list more, and a search step swaps the right ones in mid-conversation with new keyterms and a new transcription prompt. I ran the converter over the whole official registry. 12,012 servers answered without a login, listing 202,191 tools, and every tool converted. Then I gave 30 of them one spoken question each, untuned. 29 sessions ran, 14 answered from live data, and when a tool ran the median was 3.5 seconds voice to voice. A later A/B moved every session to the fastest transcription mode: 0.6 seconds quicker at the median, with nothing heard worse. Speech is bad at identifiers. Before I added a gate, a misheard wallet address reached the server 5 times, and a right hearing looks exactly like a wrong one. Now an address only runs if it comes from the paste box or an earlier tool result, tools that change state wait for a yes, and a paid tool says its x402 price instead of failing. The LLM Gateway writes starter questions for each server. The demo video was recorded live, and every voice figure comes with session IDs from Session History. Two places where the API behaves differently from its docs are written up with scripts that reproduce them. MIT licensed and live now.