Benchmarking affaan-m/ECC in Cursor: Agent Harness Rules Without Context Exhaustion

affaan-m/ECC is a standardized operating system for agents, combining skills, instincts and execution policies in a single harness. However, when using ECC with Cursor, context bloat can be a problem. When you add many rules, language-specific settings and MCP tool schemas to your agent's system prompt, it uses up tokens quickly. This slows down the model's reasoning abilities and increases the time it takes for each turn to complete in long multi-turn refactoring loops.

To set up ECC inside Cursor, you need to modularize it. Instead of putting all ECC skills into Cursor's main prompt, install the repository and only link the needed language rules to the .cursor/rules/ folder. Clone the ECC repository in the ~/.config/ecc directory, then create a .cursor/rules/ directory and copy the common and TypeScript rules from the ECC repository into it. This keeps your context lightweight, injecting language rules only when necessary.

Next, you need to set up the Modular Multi-Purpose Communication (MCP) tool configurations. ECC provides modular MCP servers for test runners and filesystem inspections. Add these servers to your project's .cursor/mcp.json file by specifying the command and arguments for the npx package that runs @ecc/harness-mcp.

Finally, to reduce the impact of multi-turn context burn, switch Cursor to a fast proxy endpoint like B-Lost's. B-Lost provides native prompt caching at the gateway layer, which automatically caches static harness rules and accumulated turn history. This cuts heavy multi-turn context costs by 80-90% without losing chat history or degrading reasoning.

Add the following to your ~/.cursor/settings.json file to enable this change: { "cursor.customModel.override": true, "cursor.customModel.baseUrl": "https://api.b-lost.com/v1", "cursor.customModel.model": "claude-3-7-sonnet-20250219" }.