FarHand

Gives a coding agent hands on a remote machine and keeps its hands off the local one. https://farhand.cogflux.io

FarHand is an MCP server. Point OpenCode, Claude Code or Codex at it, and every command the model runs and every file it reads or writes happens on a remote host over your existing SSH setup. The local machine stays closed, with one exception you control: a list of local folders the model may look into and upload from. Nothing that looks like a credential is ever sent to the remote, and everything that happens is written to a local audit log.

┌─ OpenCode ─┐   ┌─ Claude Code ─┐   ┌─ CogFlux (later) ─┐
│ thin plugin │   │ .mcp.json     │   │ --mcp-config      │
└──────┬──────┘   └──────┬────────┘   └─────────┬─────────┘
       └─────────────────┴── MCP over stdio ────┘
                       farhand (one binary)
        ┌──────────────────┼───────────────────┐
   remote tools       local tools          audit log
   remote_shell        local_ls / local_read     one JSON line per action
   remote_read/write  upload / download
   remote_edit/ls     (allowlisted folders only)
   remote_glob/grep
        │
   your OpenSSH: ~/.ssh/config, keys, agent, jump hosts — one multiplexed session

Why not opencode serve on the remote?

That works, but the agent's provider keys and configuration have to live on the remote, and every agent needs its own way of doing it. FarHand keeps the agent — and its credentials — local, and gives every MCP-capable agent the same remote hands.

Install

Local: macOS or Linux with OpenSSH. Remote: Linux with bash, timeout (coreutils), base64, an SFTP subsystem (default in sshd), and ideally ripgrep for remote_glob / remote_grep (both fall back to find / grep). Windows is not supported: FarHand relies on OpenSSH connection multiplexing over a Unix socket.

Prebuilt binary (no Rust needed) — one file into ~/.local/bin, checksum verified:

curl -fsSL https://raw.githubusercontent.com/CogFlux/farhand/main/install.sh | sh

Or download farhand-<version>-<target>.tar.gz from the releases page yourself, or cargo binstall farhand.

From source (Rust stable): cargo install --path crates/farhand puts farhand in ~/.cargo/bin; or cargo build --release and point FARHAND_BIN at target/release/farhand.

Then, either way:

farhand install opencode                  # and/or claude-code, codex — once per machine
cd ~/my-project
farhand init > .farhand.toml              # edit host, workdir, allowed_dirs
farhand check                             # connects once and reports

That is all: a project is remote exactly when it has a .farhand.toml. There is no global configuration to write unless you want one (see Configure).

See docs/USAGE.md for the step-by-step guide and per-agent configuration.

Configure

farhand init prints a commented template. The essentials:

[remote]
host = "devbox"            # an alias from ~/.ssh/config, or user@host
workdir = "~/proj"         # commands start here; relative paths resolve here; created if missing

[local]
allowed_dirs = ["."]       # the only local folders the model may see; "." = this folder

With "." the folder holding .farhand.toml is the project's local pocket: drop a file in it and the model can upload it, and download lands there. Nothing else lives in that folder — the code is on the remote — so this is the natural default. Absolute paths and ~/... work too; relative entries are only accepted in a project .farhand.toml, never .., / or ~.

The remote may be Linux, macOS, BSD — or Windows with OpenSSH Server, where commands run in PowerShell (os = "windows" under [remote] tells the model so up front; the platform is detected on connect regardless). Directory creation and tree walking go over SFTP, so uploads, downloads, remote_write and remote_edit behave the same everywhere. Transfers are streamed with many requests in flight, so a large file moves at link speed and memory stays flat whatever its size; the default cap of 1 GiB per transfer is max_transfer_bytes under [limits].

Config is found in this order: --config PATH, $FARHAND_CONFIG, ./.farhand.toml, ~/.config/farhand/config.toml. The per-project .farhand.toml is the normal one: it is what makes a directory remote. The global file is optional and serves two purposes: activation = "always" turns every directory on the machine remote using its [remote], and it is a convenient target for --config. Without any config at all, FarHand is simply inactive: the agent keeps its local tools and the server offers none.

Agents

farhand install <agent> does the wiring for opencode, claude-code and codex (--scope user|project); farhand uninstall reverses it and farhand status reports it. By default only directories that carry a .farhand.toml go remote (activation = "project"); everywhere else the agent keeps its local tools. activation = "always" in the global config makes every session remote. What each agent gets is in docs/USAGE.md.

OpenCode

farhand install opencode writes plugins/opencode/farhand.ts into ~/.config/opencode/plugins/ (or symlink it there yourself for development). The one file carries both plugin APIs, so it works with OpenCode 1.18.29+ and OpenCode 2 alike and survives an upgrade between them. The plugin registers the MCP server, removes OpenCode's local tools (bash/shell, read, write, edit, patch, glob, grep, list, the LSP tools, and in OpenCode 2 the browser_* tools and Code Mode's execute, since a local browser and execute's fetch can both open file:// URLs), refuses them again at call time as a second fence, and tells the model in the system prompt that it is working remotely.

Approval is FarHand's knob, not the agent's: [approval] mode = "ask" (default) makes the agent prompt before remote_shell, remote_write, remote_edit, upload and download and lets read-only tools run; "auto" prompts for nothing, with the audit log as the record; "strict" prompts for everything, reads included; tools = { remote_write = "auto" } overrides per tool. The plugin translates this into OpenCode's permission system at startup.

Without the plugin, the same effect comes from opencode.jsonc (OpenCode 1 format shown; OpenCode 2 reads it too, but cannot remove browser_* from config alone — use the plugin there):

{
  "mcp": {
    "farhand": { "type": "local", "command": ["farhand", "serve"], "enabled": true, "timeout": 60000 }
  },
  "tools": { "bash": false, "read": false, "write": false, "edit": false,
             "glob": false, "grep": false, "list": false, "patch": false },
  "permission": { "farhand_remote_shell": "ask", "farhand_remote_write": "ask",
                  "farhand_remote_edit": "ask", "farhand_upload": "ask", "farhand_download": "ask" }
}

The tools then appear as farhand_remote_shell, farhand_upload, and so on.

Tools

What never leaves this machine

Every byte headed for the remote — command text, file content, edits, uploads — passes the secret guard first:

• by path: .env*, *.pem, *.key, id_*, .netrc, .npmrc, credentials*, *service-account*.json, *.tfstate, agent configs (opencode.json*, .mcp.json, claude_desktop_config.json, ...), and anything under .ssh, .aws, .gnupg, .kube, .docker, .config, .claude, .opencode, .codex, ...;
• by content: private-key blocks, sk-…, ghp_… / github_pat_…, AKIA…, AIza…, xox…, Stripe, JWTs, Authorization: Bearer …, URLs with embedded passwords, and api_key = "<24+ random chars>" style assignments.

Add your own with [guard] deny_globs / deny_content. A refusal comes back to the model as a tool error naming the rule, never the matched text, and is audited the same way. There is no override flag.

The local environment is never forwarded to the remote. Commands run in the remote user's own login shell.

Threat model

FarHand is built so that the model can be wrong or hostile and the damage stays on the remote. What holds regardless of which model is driving:

• No local execution. The only process FarHand ever starts locally is ssh to the configured host (plus ssh -G to print its address). No model-supplied string reaches a local command line.
• Local reads are the allowlist, minus credentials. local_ls, local_read and upload resolve paths through the allowlist with symlinks followed first, so a link out of the folder is refused. Files the guard would refuse to upload are hidden from local_ls and refused by local_read: what the model cannot see it cannot re-encode. / and the home directory are not accepted as allowed folders.
• Downloads cannot plant anything. Every file a download would create is resolved through the allowlist again (a symlink inside the folder cannot redirect it) and checked by name: nothing credential-shaped, and nothing an editor or agent executes or trusts on opening a folder (.vscode, .idea, .git, .husky, .envrc, .claude, .farhand.toml, CLAUDE.md, ...). A refused name aborts the whole download before the first byte is written.
• Only hosts you have already met. SSH host keys are checked strictly: a .farhand.toml that arrives inside a cloned repository cannot point the agent at a machine you never connected to. Connect once by hand first.
• Everything is on record. Every call, including refusals, is one line in the audit log.

What FarHand does not do: it cannot stop a model that has legitimately read a non-secret file in the allowlist from sending it to the remote (that is the point of the allowlist), it does not restrict the remote itself (treat the remote account as the model's sandbox and give it only what that account should have), and it does not disable other MCP servers or tools your agent has that reach the local machine. Keep secrets out of allowed_dirs; the guard is a tripwire for accidents, not a substitute for that.

There is deliberately no local curl/fetch tool: curl reads local files (-d @file, file://) and reaches local-only services (Docker sockets, IDE and browser debug ports, localhost dashboards), and an approval prompt is not a defence against a request that looks harmless. Fetch from the remote with remote_shell instead; the remote has curl too.

Audit log

~/.local/share/farhand/audit/YYYY-MM-DD.jsonl, one object per action: timestamp, session, host, tool, command or path, byte count, outcome (ok / denied / error / timeout), exit code, duration. File contents are never logged, and a command the guard refused is logged as redacted.

Layout

crates/farhand-core   config, guard, local allowlist, remote (ssh + sftp), transfer, audit
crates/farhand    the `farhand` binary: MCP tools over stdio
crates/farhand/plugins/opencode   the OpenCode entry point (embedded in the binary)
packages/npm/farhand  the npm wrapper (`npx farhand`), fetches the release binary

Claude Code

farhand install claude-code registers the server and adds a PreToolUse hook (farhand hook claude-code) that refuses local tools wherever FarHand is active and answers mcp__farhand__* calls from [approval] on every call; project scope also writes a static permissions.deny. Tested at project scope.

Codex

farhand install codex --scope project registers the server, turns Codex's shell tool off, sets the sandbox read-only and translates [approval] into per-tool approval_mode. User scope registers the server only.

Roadmap

• CogFlux integration: launched via --mcp-config, audit log as an ingest source.
• Several remotes per config with a host argument on the tools.