Pi Agent Just Added CODEMODE… and It’s INSANE!

Pi Coding Agent’s MCP Update: Code Mode and the Pragmatic Shift from “No MCP”

This video analyzes Pi Coding Agent’s controversial update that introduces native MCP (Model Context Protocol) support alongside a new code mode feature. The creator, Kinto, explains why Pi’s founders—Mario Zechner and Flask creator Armin Ronacher—previously rejected MCP as “token-wasting bloat,” and how the new architecture aims to solve the very inefficiencies they criticized. Through hands-on tests and community reactions, the video evaluates whether Pi’s embrace of MCP is a necessary compromise or a betrayal of its minimalist roots.

🧠 The Core Problem: Tool Calls Flooding the Context Window

Traditional AI agents handle tool calls sequentially: the model requests data, receives a raw payload, reads it in its context window, then requests the next piece. This back-and-forth quickly clogs the model’s memory with bulky intermediate data. The video likens it to “hiring a master chef but forcing them to inspect every potato one by one.” The result is high token usage, increased cost, and frequent truncations—especially when processing many items (e.g., 50 support tickets or 100 GitHub issues).

⚡ Code Mode: A JavaScript Sandbox for Efficient Orchestration

Pi’s code mode addresses this by letting the AI write a small JavaScript script that runs inside Pi’s embedded QuickJS sandbox. Instead of the model acting as middleman for every tool call, the script can:

• Loop through data, call tools in parallel, perform calculations, remove duplicates, and filter irrelevant results.
• Keep bulky intermediate data inside the sandbox, returning only a clean summary (e.g., a table of matching tickets or critical issues).
• Drastically reduce context load, token consumption, and cost.

🔍 MCP Integration with Progressive Disclosure

Pi now exposes MCP tools as callable JavaScript functions inside code mode. To avoid dumping every tool’s description into context, Pi uses tool search and progressive disclosure:

• The model receives only a lightweight overview of tool categories.
• When needed, it uses search_tools and describe_tool functions to fetch specific tool details.
• This prevents thousands of tokens from being wasted on tools the model never uses.

Additionally, code mode can call system‑one models (e.g., Jeff or WeKnow) directly from the JavaScript loop. For example, after fetching 200 GitHub comments via MCP, the script can pass each comment through a classifier (e.g., “is this user frustrated?”) before the frontier model sees only the final aggregated result.

📊 Performance Tests: Plain MCP vs. Code Mode

Two tests were conducted using Pi with GPT‑5.6 (salt set to high):

1. NHTSA Vehicle Safety API – Comparing 10 SUVs by recalls, complaints, crash ratings, and investigations.

   • Plain MCP: Only 4 of 10 completed; 95,000 cache tokens; cost $0.26.
   • Code Mode: All 10 completed with zero truncations; cache tokens dropped to 21,000; cost $0.09.

2. 100 Public VS Code Issues (September) – Classifying and analyzing issues.

   • Plain MCP: Stopped at 50 due to context overflow; uncache tokens ~92,000; cost $0.35.
   • Code Mode + local classifier (Wino 12B): Completed all 100; uncache tokens ~30,000 (67% less); cost $0.14.

The results demonstrate massive token and cost savings while maintaining or improving task completion.

💬 Community Reactions: Pragmatism vs. Purity

The update sparked mixed reactions:

• Enterprise advocates argued that Pi’s anti-MCP stance ignored real‑world needs like remote credentials, security boundaries, and tool management across teams.
• Middle‑ground voices compared MCP to USB‑C or HDMI—flawed but necessary for ecosystem unity.
• Minimalist Pi users worried about bloat, with one commenter saying, “So Pi also accruing cruft now.”
• Armin Ronacher defended the change, stating “Nothing is loaded by default that was not loaded before,” and clarified that code mode is not “bash with extra steps” but a way to orchestrate tools inside the harness.

The video notes that other tools (Cloud Code, CodeDex, Open Code) are adopting similar lazy tool loading or running MCP through their own code mode, signaling an industry trend.

Key Takeaway

Pi’s shift from “no MCP” to native support with code mode represents a pragmatic compromise: it retains the efficiency of sandboxed processing while embracing the dominant ecosystem protocol. The tests confirm that code mode can dramatically reduce token usage and cost, but the community remains divided on whether this adds unnecessary complexity. Ultimately, the update shows that engineering purity often yields to ecosystem momentum—and that smart architecture can mitigate the downsides of a universal standard.