Copilot Agent-Mode Cost Cut by 60%

Token Proctor has been released as a local layer sitting on top of GitHub Copilot and token-billed LLMs.

• It validates prompts on a 0–100 score, fills in missing parts, and recommends a model suited to the task type.
• Cost is pre-estimated as tokens × turns × $, and it also shows the proportion consumed relative to a monthly plan.
• Optimization criteria are split into three: tokens, turns, and balanced, with turn-heavy work like agent loops routed by the turns criterion.

The setup comes in two forms.

• VS Code chat participant (@proctor): based on vscode.chat and vscode.lm.
• MCP server: the same core can be used across Copilot CLI, Copilot agent mode, Claude Desktop, Cursor, and more.

v0.4 adds turn-aware cost projection. Using a 20-turn refactoring task as an example, it explains that while Claude Sonnet 4 could cost 20 premium-equivalent requests, o4-mini brings that down to around 6.6, a reduction of about 67%.

It also applies accurate token counting by default via js-tiktoken, and manages allowed/blocked models, redaction rules, audit logs, plan limits, and LLM judge settings through the .token-proctor.json policy file.