Pi adds MCP support after publicly mocking it—engineering team explains what changed
bibryam · x · 2026-10-05
Pi (Earendil) once proudly declared it wouldn't support MCP and repeatedly dismissed it on podcasts—yet MCP is now core functionality. The engineering team explains why.
- The world isn't static: MCP improved significantly over the past year, though that alone wasn't enough
- MCP existed as an extension; it moved into core because the underlying changes were generally useful—e.g., they make it easier to run Jev inside Pi. What Pi needs resembles what MCP needs: a sandbox in the form of an interpreter
- Biggest remaining flaw: MCP is still hard to compose. Even with codemode, the problem lies less in MCP itself than in the server ecosystem and harnesses built around dumping tools into context
A representative case of the agent tooling ecosystem's shift on MCP.
More from coding & agent
- Dev tags Render's CEO over rate-limits bottlenecking agent debugging — DanielLockyer · 2026-10-05
- Dev: Rate-limits are now the biggest bottleneck to agent debugging — DanielLockyer · 2026-10-05
- 349K lines of code, 343K lines of tests: how a healthcare AI team keeps validation in pace — alex_verem · 2026-10-05
- DBS Framework: open-source reusable AI skills across Claude, ChatGPT, and Codex — Yasirres · 2026-10-05
- Dev building own MCP hits client inconsistencies, tempted by the very 'if clientname ==' hacks everyone mocks — lucasmeijer · 2026-10-05
- Meta's open-source Kats packs forecasting, anomaly detection and 65 time series features into one Python toolkit — mdancho84 · 2026-10-05