🐒 monkeyproof

Because letting the monkeys ssh into prod is how civilizations end.

Remote coding agent orchestration server. Spawn, stream, and interact with AI coding sessions via REST + WebSocket.

What

A lightweight process manager that lets an AI orchestrator (or you, monkey) spawn coding agent sessions on a remote box and interact with them in real-time via WebSocket.

Not a chatbot UI. Not a Claude Code wrapper. This is infrastructure -- a clean API for spawning, monitoring, and controlling CLI-based coding agents from anywhere on your network.

Why

• Your orchestrator AI lives on a laptop with 8GB RAM
• Your coding agents need 16GB+ and hours of runtime
• You need real-time output streaming, not polling SSH sessions
• You want to track sessions, replay output, and kill runaway agents
• You're tired of nohup claude --print ... & and tail -f /tmp/agent.log

Stack

• Runtime: Bun
• HTTP: Hono
• Transport: WebSocket (native Bun)
• Auth: Bearer token
• Storage: In-memory (sessions are ephemeral)

Security

MonkeyProof executes commands on the host where it runs. Bind it only on trusted networks, always set a strong AGENT_WS_TOKEN, and treat built-in agent presets as privileged because Claude presets include --permission-mode bypassPermissions for unattended work. Use direct exec with explicit commands when that permission mode is not appropriate.

API

POST   /sessions              Spawn a new coding session
GET    /sessions              List all sessions
GET    /sessions/:id          Session detail + recent output
DELETE /sessions/:id          Kill a session
POST   /sessions/:id/input    Send stdin to a running session

WS     /sessions/:id/ws       Real-time output stream + input

Quick Start

bun install
AGENT_WS_TOKEN=YOUR_STRONG_SECRET bun run src/index.ts

All HTTP endpoints require bearer authentication:

-H "Authorization: Bearer YOUR_STRONG_SECRET"

Spawn a direct exec session (default)

POST /sessions defaults to direct shell execution (type: "exec"). MonkeyProof runs the full task with /bin/sh -lc unless you pass an explicit command/args. This is the preferred path for normal automation and maintenance work because MonkeyProof executes the command directly instead of routing through Claude middleware.

curl -X POST http://localhost:3200/sessions \
  -H "Authorization: Bearer YOUR_STRONG_SECRET" \
  -H "Content-Type: application/json" \
  -d '{
    "task": "bun test",
    "cwd": "/home/skippy/Development/king/king-dashboard"
  }'

You can also be explicit:

{ "type": "exec", "task": "scripts/maintenance --dry-run", "cwd": "/path/to/repo" }

mode is accepted as a backwards-compatible alias for type. Presets are not valid for exec sessions; set type: "print" or type: "interactive" when using preset.

Spawn an agent print session

Claude/Codex print sessions are still supported, but they must be requested explicitly with type: "print" (or mode: "print").

curl -X POST http://localhost:3200/sessions \
  -H "Authorization: Bearer YOUR_STRONG_SECRET" \
  -H "Content-Type: application/json" \
  -d '{
    "type": "print",
    "task": "Fix the bug in auth.ts",
    "cwd": "/home/skippy/Development/king/king-dashboard",
    "command": "claude",
    "args": ["--print", "--permission-mode", "bypassPermissions"],
    "maxTurns": 50
  }'

Stream output via WebSocket

const ws = new WebSocket("ws://localhost:3200/sessions/abc123/ws?token=YOUR_STRONG_SECRET");
ws.onmessage = (e) => {
  const msg = JSON.parse(e.data);
  if (msg.type === "stdout") process.stdout.write(msg.data);
  if (msg.type === "stderr") process.stderr.write(msg.data);
  if (msg.type === "exit") console.log(`Done: exit ${msg.code}`);
};
// Send stdin
ws.send(JSON.stringify({ type: "stdin", data: "yes\n" }));

Configuration

Env Var	Default	Description
PORT	3200	Server port
AGENT_WS_TOKEN	CHANGE_ME_AGENT_WS_TOKEN	Bearer auth token. Replace before exposing the service; the placeholder is intentionally unsafe.
MAX_SESSIONS	50	Max concurrent sessions
OUTPUT_BUFFER_SIZE	2000	Lines of output to buffer per session
SESSION_TTL_MS	3600000	Auto-cleanup exited sessions after (ms)
INTERACTIVE_SESSION_TTL_MS	7200000	Auto-cleanup exited interactive tmux sessions after (ms)

Agent Presets

Instead of specifying command + args every time for print/interactive agent sessions, use presets and set type/mode explicitly:

curl -X POST http://localhost:3200/sessions \
  -H "Authorization: Bearer YOUR_STRONG_SECRET" \
  -H "Content-Type: application/json" \
  -d '{"type": "print", "task": "Fix the bug", "cwd": "/path/to/repo", "preset": "claude"}'

Built-in presets: claude, claude-sonnet, claude-opus, claude-interactive, claude-interactive-sonnet, claude-interactive-opus, codex, codex-auto

License

MIT -- go nuts, monkeys.