Agent Session Manager
kjy7097/agent-session-manager
Painel unificado para sessões Claude Code & Codex em múltiplas máquinas
Ver no GitHubCentro de controle web auto-hospedado para gerenciar sessões de Claude Code e OpenAI Codex em múltiplas máquinas via SSH. Centraliza abertura, busca, retomada e sincronização de sessões de coding-agents através de um único navegador, com suporte a tunelamento SSH automático e sincronização de skills.
Instalação
npx convez-ai kjy7097/agent-session-manager1 comando no terminal: detecta o formato do repositório (marketplace ou plugin avulso) e instala via Claude Code. Requer Node 18+ e o Claude Code instalado.
Métricas
Casos de uso
- Gerenciar sessões de coding-agents distribuídas em múltiplos computadores
- Sincronizar skills e configurações entre máquinas automaticamente
- Retomar, fazer fork ou alternar entre sessões Claude e Codex em um único painel
Veredito

Solução robusta para desenvolvedores que usam múltiplas máquinas com Claude Code e Codex. Exige conhecimento de SSH e infraestrutura, mas elimina fricção operacional de gerenciar agentes distribuídos. Produção-ready com suporte multiplataforma.
README
A tradução para o português deste README ainda está na fila.
Agent Session Manager
English | 한국어
A self-hosted web control center for CLI coding-agent sessions across multiple
machines. It drives Claude Code (claude --remote-control) and OpenAI
Codex behind one control plane — folders and session lists merge both agents,
with per-agent actions on each session.
Open a remote session in any folder on any of your computers, browse / search /
resume / delete past sessions per folder, and keep your custom skills in sync —
all from one browser tab. No more SSHing into each box and typing
claude --remote-control by hand.
Works over any SSH-reachable network — LAN, a VPN/mesh (e.g. Tailscale), or public IPs. There is no dependency on Tailscale; a machine is just a host + SSH user.
Features
- Multi-machine — register machines by host + SSH user (Windows and Linux targets); reach each agent over an auto-reconnecting SSH tunnel.
- Multi-agent — Claude Code and OpenAI Codex sessions side by side (🟧 / 🟢 badges), one unified "+ New session" button — then pick the agent.
- Claude sessions open in a
claude.ai/codetab (bridge URL); resume / fork / stop / delete. - Codex sessions run through the Codex app-server so they also appear in the Codex desktop app (see below). Continue a session in the in-app chat panel or in an interactive terminal, preview the transcript, delete.
- Terminal hand-off — open a codex TUI on the PC your browser is on
(the control plane detects the client machine by its tailnet IP; remote targets
are reached with
ssh -t). - Model selection — choose a model per new session / per resume, and set a server-saved default from the header.
- Rewind & cross-agent switch — branch a new session from any point in a transcript, or continue the conversation in the other agent (🟧 Claude ↔ 🟢 Codex).
- Browse by folder, per machine; search sessions by title and conversation content; preview a transcript before resuming.
- Automatic skill sync — your
~/.claude/skillsconverge to the newest version of each skill across all machines (last-write-wins by mtime), excluding credential/encrypted files. Edit anywhere; it propagates. - Survives reboots — agents run as OS services (Windows Task Scheduler / Linux systemd / macOS launchd); the control plane + tunnels run under launchd.
- Korean/English UI — 🌐 toggle; the preference persists server-side.
- Single static web UI (vanilla JS), mobile-friendly. Pure Python 3 standard
library on the backend (plus
pywinptyon Windows agents).
Codex desktop app integration
ASM creates Codex sessions through the Codex app-server (thread/start +
turn/start), so they are tagged source=vscode and show up in the Codex
desktop app's project list — a codex exec session would be source=exec and
the app hides it. The app picks new sessions up on its next poll (no reconnect
needed).
One thing ASM cannot do: add a project/folder to the desktop app. The app keeps its project list internally (not in any file ASM can write), so:
- Add the folder once in the Codex desktop app (as a project).
- ASM then lists that folder under
/codex/projectsand creates sessions in it — and those sessions appear in the app under that project.
In other words: projects are added in the app (one-time, manual); sessions are created by ASM (unlimited, automatic). There is no "open in app" button for Codex — the app can't be driven from outside — but sessions ASM creates simply appear there.
Architecture
browser ──► control plane (this host) ──► SSH tunnel ──► agent (each machine)
registry + UI + auth proxy ~/.claude + ~/.codex
- agent (
csm/agent.py) — one per machine. Reads that machine's~/.claudeand~/.codex(folders, sessions, transcripts, skills), launches detachedclaude --remote-controlsessions in a real PTY (POSIXpty.fork, or ConPTY viapywinptyon Windows), and opens codex terminals on the machine's desktop. Bound to127.0.0.1; bearer-token (HMAC) auth. - codex adapter (
csm/codex.py) — parses Codex rollout JSONLs, drives the codex app-server (thread/start/turn/start/thread/resume) so sessions are app-visible, and readsconfig.toml [projects]. - control plane (
csm/control.py) — serves the web UI, holds the machine registry (fromcsm.config.json), proxies browser requests to each agent over its local SSH-tunnel port (attaching the per-machine secret), and routes the terminal hand-off to the right machine by client IP. - tunnels (
csm/tunnels.py) — keeps a forward-only SSH tunnel (localPort → host:agentPort) alive per machine. - skill sync (
csm/skillsync.py) — periodic per-skill convergence across machines.
Requirements
- Control host: Python 3.10+, OpenSSH client, an SSH key.
- Each machine: Claude Code CLI installed and logged in (
claude auth), Python 3, OpenSSH server, and the control host's SSH public key authorized. For Codex features: the Codex CLI installed and signed in (ChatGPT login or a local provider) — optional per machine. - Windows agents additionally use
pywinpty(installed automatically by the deploy script).
Quick start
git clone <this-repo> && cd agent-session-manager
./run.sh # starts the local agent + control plane
# open http://127.0.0.1:8765
On first run a csm.config.json is created (random secret). See
csm.config.example.json for the schema. To expose it to your other devices,
bind the control plane to your VPN/LAN IP in csm.config.json (control.host).
Add a machine
- Authorize the control host's SSH key on the target (one-time), with the
enroll script served by the control plane:
- Windows (elevated PowerShell):
irm http://<control-host>:8765/enroll | iex - Linux (terminal; sudo will prompt):
curl -fsSL http://<control-host>:8765/enroll?os=linux | bash
- Windows (elevated PowerShell):
- From the control host, deploy the agent:
The target OS is auto-detected; the agent autostarts via a Windows scheduled task or a Linux systemd user service, a tunnel opens, and the machine lands inpython scripts/deploy_agent.py <id> <host> <ssh-user>csm.config.json. Refresh the UI — it appears. (Or use the Add machine dialog in the UI, which does the same.)
Security
- Agents bind to
127.0.0.1and are reached only via SSH tunnels; each agent requires a per-machine bearer secret. The web UI/secret never leave the control host. Keep the control plane on a trusted network (localhost/VPN/LAN) — an agent can launch processes in arbitrary folders, so treat access as you would shell access. csm.config.jsonholds secrets and is git-ignored. Never commit it.
Changelog
See CHANGELOG.md. Current version: 0.5.0.
License
MIT — see LICENSE.