IM.codes

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, GitHub Copilot, Cursor, OpenCode, Qwen, OpenClaw, and more. 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.

imcodes bind https://app.im.codes/bind/<key>

bound to app.im.codes · daemon started · registered as system service

Self-hosting is strongly recommended. app.im.codes is a shared test instance with no uptime guarantees — it may be rate-limited, targeted, or unavailable. This is a personal project with no commercial support. For anything beyond evaluation, deploy on your own infrastructure.

screenshots

Desktop

iPad / Tablet

Mobile

Apple Watch

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

why

When you leave your desk, most coding-agent workflows fall apart. The agent is still running in a terminal, but continuing usually means SSH, tmux attach, remote desktop, or waiting for your laptop. That is only half the problem. Hard agent work also needs better judgment: one model can fall into patterns, miss issues, or lose context when you switch providers. IM.codes pairs always-reachable sessions with shared memory for summarized completed work and structured Team cross-model review, so multiple agents can challenge plans before code lands. It does not make output perfect, but it reduces single-model blind spots while keeping the work continuous.

shared agent context

IM.codes keeps a searchable memory of solved work — structured problem→solution summaries instead of raw prompt logs. Relevant history is recalled automatically for new prompts and session startup, and every injection stays visible in the timeline.

Personal memory stays local by default, with optional cloud sync for processed summaries across your devices.
Enterprise shared context lets teams reuse fixes across workspaces and projects with queryable memory, stats, and inspection UI. It is still under active development and has not been fully production-tested yet.
Multilingual recall uses local embeddings plus pgvector-backed server search, so related fixes can be found across different languages.
Timeline cards show exactly what was injected, with relevance score, recall count, and last-used metadata.

managed MCP tools

Supported SDK-backed agents receive a daemon-managed stdio MCP server with one runtime-scoped surface for memory, agent messaging, and scheduled follow-ups, without raw auth tokens or ad hoc shell commands.

search_memory recalls prior work, decisions, preferences, bugs, commits, deployments, and discussed context, returning compact summaries plus projectionId values.
get_memory_sources expands a relevant memory hit into provenance snippets when exact prior instructions, bug details, or source evidence matter.
save_observation and save_preference persist useful facts and stable user preferences.
send_list_targets and send_message let agents discover sibling sessions and send scoped text, file references, replies, or broadcasts.
cron_create, cron_list, cron_update, and cron_delete manage future structured sends for reminders, recurring checks, delegated reviews, and follow-ups.
Runtime-bound identity prevents forged namespace, user, server, token, or routing fields; provider-level ready/degraded status and redacted recent calls stay visible in the UI.

supervised execution

IM.codes can supervise supported transport sessions turn by turn instead of relying on blind auto-continue. Auto checks whether the latest turn is done, should continue, or should return control to you.

Per-session Auto modes: Off, Supervised, and Supervised + audit.
At the idle boundary, Auto can classify a turn as complete, continue, or ask_human, then dispatch the next continue prompt in the same session.
Supervised + audit can run an audit→rework loop before handing control back, so finished work is reviewed instead of trusted blindly.
Global defaults seed new sessions. Set a default supervisor backend, model, and timeout once — each session snapshots them at enable time and can still override backend/model/timeout and audit mode individually.
Two-layer custom supervision instructions. Keep a global supervision persona plus a per-session addition — by default the two are concatenated into the prompt, or tick the session's “Override global” checkbox to ignore the global value for that session. Editing the global persona takes effect on already-enabled sessions on the next turn without re-enabling Auto.

what it is not

Not another AI IDE
Not just a chat wrapper
Not just a remote terminal client
Not a replacement for Claude Code, Codex, Gemini CLI, OpenClaw, or Qwen
It is the messaging/control layer around them

features

shared agent context

problem→solution memory, multilingual recall, automatic per-message and startup injection with visible timeline cards

managed MCP tools

ten runtime-scoped tools for memory search/sources, preferences, scoped Send, and Cron scheduling

supervised execution

turn-level completion checks, auto-continue, and optional audit→rework loops for transport-backed agents

remote terminal

browser & mobile, no SSH/VPN, real-time PTY streaming at 12fps

file browser & git

tree view, upload/download, +/- stats, safe HTML quick preview, inline image previews with lightbox zoom

local web preview

preview localhost from any device via secure tunnel, supports HMR

mobile, watch & notifications

biometric auth, push notifications, shell keyboard input, watch quick replies

cross-provider audit

multi-agent Team discussions with customizable phases (audit, review, brainstorm…) — different models catch different issues before code is written, reducing rework

CLI + SDK access

Access multiple agents through both CLI and SDK integrations. OpenClaw and Qwen also stream natively with real-time deltas, tool tracking, and session restore.

terminal + chat

raw CLI or structured view with parsed tool calls

Discord-style sidebar

server icons, session tree, unread badges, idle flash

@ picker

@files, @@agents, @@all(config) for group dispatch

multi-server

manage agents across machines from one dashboard

pinnable panels

drag any window to sidebar, cross-device sync

sub-sessions

spawn parallel agents on the fly

repository dashboard

issues, PRs, branches, commits, CI/CD

scheduled tasks

cron-style automation, timezone-aware, execution history with jump-to-session

cross-device sync

tab order, pinned tabs & panels synced to server

7 languages

EN, 中文, 繁體, ES, RU, JA, KO

real-time stream

live output, zero polling, zero message limits

self-hosted

your infra, your data, one-command deploy

OTA updates

upgrade daemons from the web UI

architecture

You (browser / mobile) │ WebSocket Server (self-hosted, relays connections) │ WebSocket Daemon (your machine, manages sessions) │ tmux / transport / managed MCP Agents (multiple agents via CLI + SDK · native transport backends)

The daemon runs on your dev machine and manages process-backed sessions through tmux plus transport-backed sessions through SDKs and network protocols. It also owns the managed MCP server for runtime-scoped memory, Send, and Cron tools. The server relays connections between your devices and the daemon. Everything stays on your infrastructure.

download

Also available as a web app

⌚ iPhone · iPad · Apple Watch Watch app includes session list, unread counts, and quick replies.

install

$ npm install -g imcodes

click to copy

quick start

Use app.im.codes or self-host the server
Run imcodes bind https://app.im.codes/bind/<key>
Open the dashboard — your agents are live

openclaw connect

If OpenClaw is running on the same machine as the daemon, connect IM.codes to the local OpenClaw gateway with:

imcodes connect openclaw

Default gateway URL: ws://127.0.0.1:18789
Token auto-detected from ~/.openclaw/openclaw.json, or pass --token / OPENCLAW_GATEWAY_TOKEN
OpenClaw sessions and child sessions sync into IM.codes automatically
Saved to ~/.imcodes/openclaw.json so the daemon can reconnect automatically
Use imcodes disconnect openclaw to remove the saved config

imcodes connect openclaw --url wss://gateway.example.com
OPENCLAW_GATEWAY_TOKEN=... imcodes connect openclaw

Only tested on macOS so far. Remote plain-text ws:// URLs require --insecure.

self-host

Deploy server + daemon on a single machine. Requires Docker and a domain with DNS pointing to the server.

npm install -g imcodes
mkdir imcodes && cd imcodes
imcodes setup --domain imc.example.com

Generates all config, starts PostgreSQL + server + Caddy with automatic HTTPS, creates admin account, and binds the local daemon. Credentials printed at the end.

To connect additional machines:

npm install -g imcodes
imcodes bind https://imc.example.com/bind/<api-key>

Manual setup:

git clone https://github.com/im4codes/imcodes.git && cd imcodes
./gen-env.sh imc.example.com # generates .env, prints admin password
docker compose up -d

Generated docker-compose.yml uses pgvector/pgvector:pg18 for PostgreSQL.

supported agents

Multiple agents support both CLI and SDK integrations.

Claude Code Codex Gemini CLI OpenClaw Qwen GitHub Copilot Cursor OpenCode Shell + more

requirements

macOS, Linux, or Windows (native experimental; WSL fully supported)
Node.js ≥ 20
tmux (Linux/macOS/WSL). Windows native uses ConPTY (built-in). Multiple agents also support SDK-backed sessions.

about

Personal project. Built almost entirely by Claude Code, with contributions from Codex and Gemini CLI.

IM.codes is an independent open-source project and is not affiliated with, endorsed by, or sponsored by Anthropic, OpenAI, Google, Alibaba, OpenClaw, or any other company whose products are mentioned. All product names, trademarks, and registered trademarks are the property of their respective owners.