IM.codes
The messaging, memory, and control layer around terminal AI coding agents: take over sessions from mobile, share cross-provider agent memory, and supervise multi-agent work with audit.
Evidence shows independent node credentials, revocable execution, runtime-bound MCP identity, feature gates/kill switches, Auto Deliver never staging/committing/pushing and failing closed to a human, and lint-staged blocking likely secrets. Deducted: scoping claims are mostly README assertions; dependency security shows only pinned versions and lock verification, no audit/scan evidence.
Ack reliability (reconnect grace, dedup, timeout retry, handshake-window buffering) has concrete tests, with structured failure reasons (daemon_offline/ack_timeout). Deducted: the dependency surface is very broad (multiple vendor SDKs, native modules); availability claims are not fully supported by evidence.
Multi-provider, multi-platform, multilingual audience with explicitly stated boundaries (OpenClaw macOS-only, remote desktop Windows-only, enterprise memory unproven) and clear Node>=22 requirement. Deducted: trigger precision (Auto classification, alias resolution) is described but not statically verifiable in detail; environment/self-hosting docs are thin.
Well-organized multilingual README and a complete MIT license (full marks). Deducted: no install/self-hosting instructions in evidence, no usage examples or FAQ, no CHANGELOG (only 0.1.2 in package.), and an explicitly no-SLA personal project with no stated maintenance/update path.
Structured cards, evidence and scores make outputs usable; the unified IM/memory/audit layer offers real marginal value. Deducted: the stack requires a self-hosted server, daemon, controlled nodes and multiple AI subscriptions, and the evidence quantifies no cost/benefit.
Some claims are corroborated by tests and CI (WS-bridge reliability, auth/CSRF, Windows signing and self-upgrade rollback qualification). Deducted: many security/feature claims (memory redaction, cloud-sync isolation) cannot be traced to implementation in the provided files, and the README mixes marketing narrative with factual statements.
- Static review only: all security claims (scoping, redaction, isolation) are unverified by execution and rest on source/docs.
- Publisher identity is unverified; remote desktop and controlled execution are high-privilege capabilities — audit the implementations before deploying.
- Very large dependency surface with no evidence of vulnerability scanning; run dependency audits before use.
- The project disclaims SLA and stability and is at version 0.1.2; APIs may change without notice.
- Install and self-hosting documentation is missing from the evidence; real deployment cost may exceed what the README implies.
What does this agent do, and when should you use it?
IM.codes (repository im4codes/imcodes) is a self-hosted messaging and control platform built around terminal coding agents such as Claude Code, Codex, Gemini CLI, OpenClaw, and Qwen. It consists of a self-hosted server, a daemon running on your dev machine that manages sessions via tmux or per-provider SDKs, and browser/mobile/web clients. Core capabilities include remote terminal plus file/Git views, controlled-node remote desktop, shared agent memory (summarized completed work with multilingual pgvector recall), a daemon-managed MCP tool surface for memory, agent messaging and Cron scheduling, and multi-agent discussions with cross-provider audit. OpenSpec Auto Deliver drives a change folder through an end-to-end supervised pipeline with audits, module scoring, and rework gates. Everything runs on your own infrastructure — server, daemon, and data are never delegated to third-party hosting.
After installing, imcodes bind <server-bind-url> registers your machine as a daemon; the daemon manages agent sessions through tmux or transport-backed SDKs (Claude Code SDK, Codex SDK, OpenClaw gateway, Qwen). From a browser or phone you can open the raw terminal or a structured chat view, browse the file tree, see per-file Git add/delete line counts, and preview localhost through a secure WebSocket tunnel. Agents receive a ten-tool runtime-bound MCP surface: search_memory, get_memory_sources, save_observation, save_preference, send_list_targets, send_message, and cron_create/cron_list/cron_update/cron_delete. Agents coordinate with imcodes send "target" "message" [--reply|--all], protected by depth, rate, and broadcast-cap circuit breakers. Auto supervision classifies each completed turn as complete/continue/ask_human at the idle boundary, and supervised_audit adds an audit-then-rework loop. Team discussions dispatch multi-agent rounds via @@all(config) using modes like audit/review/discuss/brainstorm. Controlled Nodes support scoped commands, single-file transfer, and typed Computer Use; capable Windows nodes add browser-based remote desktop.
- A developer away from their desk: continue watching terminal output, reviewing file changes, and previewing localhost from phone or tablet without SSH or port forwarding.
- An engineer with multiple machines: enroll several computers as Controlled Nodes and let one agent run scoped commands and file transfers across all of them.
- A team wary of single-model output: use Team's audit mode to have Claude Code, Codex, Gemini CLI, and others cross-review a plan before any code is written.
- An OpenSpec user: hand a change folder to Auto Deliver for automated spec audit, implementation, validation, scoring, and rework gates.
- Automation and on-call scenarios: schedule Cron-triggered agent checks, or run a Python script session that tails a log and calls
imcodes sendto trigger a fix when an ERROR appears.
What are this agent's strengths and limitations?
- Cross-provider memory layer: only final assistant.text outputs are materialized, embedded multilingually, and recalled via local semantic search plus server-side pgvector, with injection cards showing relevance score, recall count, and last-used time.
- Fail-closed automation: Auto supervision and Auto Deliver hand control back on timeouts, invalid output, or bad config, and Auto Deliver never stages, commits, or pushes — the user remains the final release owner.
- MCP tools are runtime-bound to the current session/project/user/server, so agents cannot forge namespace, token, or routing fields; per-provider MCP readiness and degraded states are visible in the UI.
- Self-contained deployment boundary: server, daemon, and data stay on your own infrastructure, and localhost preview tunnels through your own server with no third-party tunnels.
- The author explicitly states this is a personal project with no warranties, no SLA, and no guarantees of stability, security, or backward compatibility — use at your own risk.
- Deployment has real overhead: full self-hosting needs Docker, a domain with DNS, PostgreSQL + pgvector + Caddy; the daemon needs Node.js >= 22 and tmux on Linux/macOS.
- Feature gaps by platform: remote desktop currently works only on supported Windows controlled nodes (macOS/Linux nodes lack it), and
imcodes connect openclawhas been tested on macOS only. - Personal cloud sync of shared context and enterprise workspace memory are under active development and, per the README, not fully production-tested.
How do you install or deploy this agent?
Requirements: macOS or Linux (Windows supported experimentally via ConPTY), Node.js >= 22, tmux on Linux/macOS, and at least one supported AI coding agent. Install the CLI: npm install -g imcodes. Quick evaluation: use app.im.codes and bind your machine with imcodes bind https://app.im.codes/bind/<api-key> — this starts the daemon and registers it as a system service. For real use, self-host (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, which generates config, starts PostgreSQL (pgvector/pgvector:pg18) + server + Caddy with automatic HTTPS, creates the admin account, and binds the local daemon. Manual path: git clone https://github.com/im4codes/imcodes.git, run ./gen-env.sh imc.example.com, then docker compose up -d, and bind your dev machine with imcodes bind.
How do you use this agent?
After binding, sign in from the browser or the iOS app (iPhone/iPad/Apple Watch, plus the web app at app.im.codes), pick a server and session, and open the remote terminal or structured chat view. Type ^ to pick a Controlled Node target, @ to search project files, @@ to select Team dispatch targets, and @@all(config) to message all configured agents. Agent-to-agent examples: imcodes send "Plan" "review the changes in src/api.ts"; imcodes send "Cx" "run tests" --reply; imcodes send --all "migration complete, check your end". Set each session's Auto mode to off/supervised/supervised_audit individually. To attach a local OpenClaw gateway: imcodes connect openclaw (defaults to ws://127.0.0.1:18789, reuses the token from ~/.openclaw/openclaw.; tested on macOS only).