fwd

Move your coding project and agent session to remote compute.

fwd provisions or connects to a remote target, synchronizes the current project, prepares its toolchain, and starts a persistent shell, command, Claude Code, or Codex session in tmux. Disconnect your laptop and return later: the session, durable tasks, and logs remain available.

laptop                                  remote SSH / RunPod / Lambda / Slurm
project  ───────────── sync ──────────▶  project checkout
agent context ──────── optional ──────▶  Claude or Codex state
                                         tmux session + durable task manager
terminal ◀────────── attach / stream ──  persistent work

Install

uv tool install fwdit

Or try it without installing:

uvx --from fwdit fwd --help

Requires Python 3.12+, ssh, and rsync locally. Run fwd doctor to check target-specific requirements.

Quick start

Run fwd from the project you want to move:

fwd                              # attach to this project's session, or create one interactively
fwd up runpod codex              # launch Codex on RunPod; auto-attach in a human terminal
fwd up --detach runpod codex     # launch Codex but stay in the local terminal
fwd up --target work             # use a configured SSH/cloud/HPC target
fwd up -- python train.py        # launch and stream a durable command
fwd ls                           # inspect sessions and live status
fwd attach SESSION               # reconnect from a human terminal

Bare fwd is the interactive reuse workflow. Agent launches auto-attach in a human terminal; pass --detach to stay local. Scripts and coding-agent environments remain non-attaching. Detach from tmux with Ctrl-B D.

Need a saved target first?

fwd targets add                  # interactive target setup (also spelled fwd setup)
fwd targets add ssh --host my-box --target-name work
fwd targets ls                   # saved targets, their backend, and the default
fwd config --example runpod      # current generated config reference

fwd targets manages saved targets — ls, add, info, update, rm — while fwd ls and fwd rm manage the sessions running on them. Removing a target only edits configuration; use fwd rm to destroy compute.

Docs / User Guide

Read Getting started for the full first-session walkthrough.

Choose a target

• SSH: an existing host, direct address, or OpenSSH alias.
• RunPod: CPU or GPU Pods with per-session persistent network volumes by default.
• Lambda Cloud: GPU instances with persistent filesystems and local-only API credentials.
• Slurm: allocations launched through persistent login-node tmux on shared scratch.

See Configuration and backends for setup, storage, and lifecycle differences.

Common workflows

Run durable work

fwd send -- pytest -q
fwd send --detach -- python train.py
fwd send --ls --json
fwd send TASK_ID                 # reattach to its log
fwd send TASK_ID --stop          # cancel the task, not the session

Every command runs in remote tmux with a durable ID and log. Streaming returns the remote exit code; Ctrl-C cancels and Ctrl-B backgrounds the viewer.

Synchronize results

fwd diff                         # compare without changing either side
fwd push                         # mirror local synchronized files to remote
fwd pull outputs/                # additive download; never deletes local files

Sync honors .gitignore, .fwdignore, and configured exclusions. Upload includes .git/ for remote agent continuity; pull and diff exclude Git metadata.

Sync continuously

Opt in to keep both sides converged automatically while a session runs, using Mutagen:

fwd sync on                      # enable for this target and start it now
fwd sync status                  # live state and any conflicts
fwd sync off                     # stop syncing

Off by default and configurable per target. Toggling takes effect immediately, even mid-session. Conflicts are reported rather than resolved destructively, and .git is never continuously synced — keep using fwd push and fwd pull for repository state. See continuous synchronization.

Forward a service

fwd ports 3000                   # localhost:3000 to remote localhost:3000
fwd ports work 8080:3000
fwd ports --ls
fwd ports --close 3000

Forwards are loopback-only and persist through a managed SSH control connection.

Stop or destroy compute

fwd stop SESSION                 # stop compute and retain configured persistent storage
fwd rm SESSION                   # permanently destroy remote resources

Both commands protect a reachable dirty remote Git worktree. rm is irreversible; force flags explicitly accept possible loss. See Lifecycle safety before automating cleanup.

Coding-agent skill

fwd ships an Agent Skills-compatible workflow for Codex, Claude Code, and other supporting agents. The first interactive invocation offers to install the bundled skill. It can also be installed directly:

npx skills add Sid-MB/fwd --skill fwd -g -a codex -a claude-code

Invoke it as $fwd ... in Codex or /fwd ... in Claude Code. See Coding agents for behavior and credential guidance.

Documentation

User guide

• Getting started: installation, first launch, target setup, and the everyday workflow.
• Commands and lifecycle: durable tasks, synchronization, inspection, port forwarding, stopping, and removal.
• Configuration and backends: config layers, SSH, RunPod, Lambda Cloud, Slurm, toolchains, and defaults.
• Coding agents: Claude/Codex transfer, follow-up turns, credentials, runtime policy, and skill installation.
• Troubleshooting: diagnostics, launch recovery, dirty worktrees, sync limits, and destructive operations.
• User documentation index: the complete end-user map.

The installed CLI is the authoritative option reference:

fwd --help
fwd COMMAND --help
fwd config --example
fwd config --schema

Unix package builds can install the generated fwd(1) manual; visible subcommands have separate pages such as fwd-up(1). See manual-page generation and packaging.

The first invocation after uv tool install fwdit or an upgrade silently installs or updates these pages in the user-local man directory, so man fwd and man fwd-up work without sudo.

Developer guide

• Developer documentation index: architecture, provider notes, validation evidence, and repository map.
• Adding a target backend
• Adding a project toolchain
• Performance benchmarking
• Manual pages
• Contributing