Imcodes

IM.codes is a sprawling, ambitious project that tries to solve a real pain: when you run multiple AI coding agents, each one forgets what the other learned, and none of them have a shared view of solved work. The shared memory layer is the most compelling piece—structured problem→solution summaries instead of raw prompt logs, recalled automatically at session startup, with every injection visible in the timeline. The multilingual recall via local embeddings and pgvector is a thoughtful touch that most memory tools skip. The managed MCP tools are a practical win: ten runtime-scoped tools (search_memory, save_observation, save_preference, get_memory_sources, send_message, cron_create/list/update/delete) that let agents persist facts and schedule follow-ups without exposing raw credentials or ad hoc shell commands. That's a meaningful security improvement over letting agents run arbitrary commands. Supervised execution is where IM.codes gets interesting. Instead of blind auto-continue, it classifies each turn as complete, continue, or ask_human, and can run an audit→rework loop before handing control back. Combined with Team cross-model review, it genuinely reduces single-model blind spots—multiple agents can challenge a plan before code lands. OpenSpec Auto Deliver extends this to a full supervised delivery run with spec audit, implementation, validation, scoring, and fail-closed gates. The AI remote desktop and Controlled Nodes are a different beast—turning Windows machines into browser-controllable nodes with scoped commands and Computer Use. That's a distinct capability that most agent orchestration tools don't offer, though it widens the attack surface and needs careful scoping. Weaknesses are real and documented. The shared test instance (app.im.codes) has no uptime guarantees and may be rate-limited, targeted, or unavailable. This is a personal project with no commercial support, and the enterprise shared context feature is explicitly 'still under active development and has not been fully production-tested yet.' You'll need to self-host and handle your own DevOps. It's also a single-person passion project, so feature roadmap risk is non-trivial. Where it fits: developers who run multiple agents (Claude Code, Codex, Gemini CLI, Copilot) and want shared context, plus power users who self-host agent infrastructure and want vendor lock-in-free memory. Where it doesn't fit: teams that need SLAs, managed SaaS, or plug-and-play setups, and users who stick with one agent and don't need cross-model audit.