im4codes/imcodes

The IM for agents. Shared Agent Context & Memory, supervised execution, and cross-agent audit across AI providers.

What it solves

IM.codes provides a messaging, memory, and review layer for terminal-based coding agents. It solves the "reach problem" where agent workflows are tied to a specific terminal or machine, allowing users to monitor and interact with agents from mobile, web, or tablets. It also addresses the "single-model blind spot" by enabling multi-agent discussions and cross-provider audits to ensure higher code quality and steadier judgment.

How it works

The project implements a daemon-managed system that integrates with various AI providers (such as Claude Code, Gemini CLI, and GitHub Copilot) through a shared memory layer and a managed Model Context Protocol (MCP) tool surface. It captures completed agent work as reusable memory, which is then injected back into future sessions. It also features a supervised execution engine that can evaluate agent turns and decide whether to continue work autonomously or trigger an audit loop.

Who it’s for

Developers who use terminal-based AI coding agents and want to maintain control and accessibility across devices, as well as teams requiring structured, multi-model review processes for complex software changes.

Highlights

  • Shared Agent Memory: Converts completed work into reusable context and uses multilingual semantic search to recall relevant history across different sessions and providers.
  • OpenSpec Auto Deliver: A structured pipeline for spec-driven changes that includes proposal audits, implementation, validation, and automatic module scoring.
  • Managed MCP Tools: Provides agents with a standardized tool surface for memory recall, agent-to-agent messaging, and cron scheduling.
  • Supervised Execution: An automated supervisor that judges completed turns to determine if a task is finished or requires rework, reducing the need for manual "continue" prompts.
  • Remote Access: Full terminal access, file browsing, and local web previews via a secure WebSocket tunnel, accessible from mobile and Apple Watch.
  • Multi-Agent Collaboration: Supports "Team discussion" where multiple models audit each other's plans and implementations to reduce biases and misses.