IM.codes

English | 简体中文 | 繁體中文 | Español | Русский | 日本語 | 한국어

The IM for agents. Shared memory, managed MCP tools, supervised execution, and cross-agent audit across AI providers.

IM.codes gives coding agents one shared memory layer and one managed MCP tool surface across providers. It turns completed work into reusable context, then injects or recalls the right history in future sessions across Claude Code, Codex, Gemini CLI, GitHub Copilot, Cursor, OpenCode, OpenClaw, Qwen, and more — with terminal access, file browsing, git views, localhost preview, notifications, multi-agent workflows, and native streaming output for transport-backed agents. Built-in Auto supervision can judge completed turns, continue work autonomously, and optionally run an audit/rework loop before handing control back. Team discussion lets multiple models review and audit each other's plans and implementations — an effective way to reduce single-model misses, blind spots, and biases.

Disclaimer: This is an actively developed personal open-source project. There are no warranties, no SLA, and no guarantees of stability, security, or backward compatibility. Use at your own risk. Breaking changes may happen at any time without notice.

Breaking Changes

• PostgreSQL default image changed to pgvector/pgvector:pg18 (instead of postgres:16-alpine). New self-hosted deployments generated from the current templates use this image for multilingual vector search in shared agent memory:

  postgres:
    image: pgvector/pgvector:pg18   # was: postgres:16-alpine
  

  The pgvector extension is enabled automatically by the server migration on first startup.

Screenshots

Desktop

iPad / Tablet

Mobile

Apple Watch

Watch support covers quick session monitoring, unread counts, push notifications, and quick replies directly from the wrist.

Download

Supports iPhone, iPad, and Apple Watch. Also available as a web app.

Why

When you leave your desk, most coding-agent workflows fall apart. The agent is still running in a terminal, but continuing the work usually means SSH, tmux attach, remote desktop hacks, or waiting until you're back at your laptop.

That reach problem is only one half of it. Complex coding-agent work also needs steadier judgment: a single model can fall into familiar patterns, miss issues, or produce unstable answers on hard tasks. Switching providers can help, but without shared context it can also lose the thread.

IM.codes is built around both needs. It keeps sessions within reach from mobile or web: open the terminal, inspect files and git changes, preview localhost from another device, get notified when work finishes, and keep multiple agents moving on your own infrastructure. It also pairs Shared Agent Context & Memory with Multi-Agent Discussions & Cross-Provider Audit: durable recall comes from summarized completed work, while Team discussion is structured cross-model review before code lands. It does not make output perfect, but it reduces single-model blind spots and helps complex work converge with more review.

It is not another AI IDE or a generic remote terminal. It is the messaging, memory, and review layer around terminal-based coding agents.

This is a personal project. I haven't written any code myself — it was built almost entirely by Claude Code, with significant contributions from Codex and Gemini CLI.

Shared Agent Context & Memory

IM.codes continuously turns completed agent work into reusable memory and feeds that context back into future sessions.

• Problem → solution memory, not log spam. Only final assistant.text outputs are materialized. Streaming deltas, tool calls, and intermediate noise are excluded.
• Personal memory with optional cloud sync. Raw and processed memory always stay local; processed summaries can optionally sync to a user-scoped cloud pool shared across your devices.
• Enterprise shared context. Teams can publish reusable memory into workspace/project scopes, inspect it in the UI, query it, and see stats instead of treating context as hidden prompt text. This part is still under active development and has not been fully production-tested yet.
• Multilingual recall. Local semantic search and server-side pgvector recall use multilingual embeddings, so related fixes can be found across English, Chinese, Japanese, Korean, Spanish, Russian, and mixed-language repos.
• Automatic injection where it matters. Relevant past work is injected both per-message and at session startup, with timeline cards that show what was recalled, why, the relevance score, recall count, and last-used time.
• User-visible inspection and control. Shared Context UI separates raw events, processed summaries, cloud memory, and enterprise memory, with query, preview, archive/restore, and processing configuration controls.

Managed MCP Tools

IM.codes exposes a daemon-managed stdio MCP server to supported SDK-backed providers. Agents get one runtime-scoped tool surface for memory, agent-to-agent messaging, and scheduled follow-ups, without raw auth tokens or ad hoc shell commands.

• Memory recall and provenance. search_memory searches the caller-bound memory namespace for prior work, project history, decisions, preferences, bugs, commits, deployments, and previously discussed context. list_memory_summaries retrieves recent compact summaries without a query. Results include compact refs plus projectionId values; get_memory_sources expands a relevant hit into provenance snippets when the model needs exact prior instructions, bug details, commit/deployment context, or source evidence.
• Memory writes. save_observation stores useful facts, decisions, or implementation notes as user-private memory candidates; save_preference stores stable user preferences through the explicit preference path.
• Agent messaging. send_list_targets lists sibling sessions in the current project, and send_message sends scoped messages, optional file path references, reply requests, or broadcasts through the same guarded imcodes send pipeline.
• Cron scheduling. cron_create, cron_list, cron_update, and cron_delete manage future structured sends for reminders, recurring checks, delegated reviews, or scheduled Team follow-ups, with target/session/project fields and optional expiration/timezone data.
• Runtime-bound identity and safety. Tool calls are bound t