Approve AI agent tool calls from anywhere.

Farmer sits between your AI coding agent and your terminal, giving you a visual dashboard to approve, deny, or respond to tool calls in real time. Desktop and mobile.

Install

npm install -g @grainulation/farmer

Quick start

# 1. Install hooks (once)
farmer connect --global   # all projects, or:
farmer connect            # current project only

# 2. Start the dashboard
farmer start

# 3. Open the token URL printed to the terminal

# 4. Start claude in any project -- hooks route automatically

Features

• Desktop + mobile dashboard -- session sidebar, permission cards with syntax-highlighted code, activity feed, mobile swipe card view
• Agent-agnostic hook protocol -- Claude Code adapter ships first; write your own for other agents
• Multi-session -- manage multiple AI sessions from one dashboard
• Multi-user roles -- admin and viewer tokens with separate permissions; viewers see read-only cards with "Waiting for admin" labels, admin controls hidden
• Trust tiers -- paranoid (approve everything, overrides session-level rules), standard (auto-approve reads), autonomous (auto-approve most)
• AskUserQuestion -- deny-to-respond pattern lets you answer agent questions from the dashboard
• Security -- two-token auth (admin + viewer), HMAC-signed invite links with expiry, CSRF protection, CSP headers, audit logging
• Data persistence -- activity and messages survive server restarts
• Stale server guard -- auto-approves when no dashboard is connected (prevents CLI blocking)

CLI

farmer start [--port 9090] [--token <secret>] [--viewer-token <secret>] [--trust-proxy] [--data-dir <path>]
farmer stop
farmer status

Tokens are persisted in a JSON token file (tokens.json in the data directory). The file stores admin and viewer tokens and is backwards compatible with the legacy plain-text format.

Hook protocol

Farmer exposes four hook endpoints. All accept POST with JSON body, localhost only:

Endpoint	Purpose
/hooks/permission	Tool permission requests (blocking -- waits for approve/deny)
/hooks/activity	Tool completion events (non-blocking)
/hooks/notification	Messages, questions, agent events (non-blocking)
/hooks/lifecycle	Session start/end events
/hooks/stop	Graceful shutdown signal

Writing an adapter

To support a new AI agent, extend BaseAdapter in lib/adapters/base.js:

import { BaseAdapter } from "@grainulation/farmer/lib/adapters/base.js";

class MyAgentAdapter extends BaseAdapter {
  get name() {
    return "My Agent";
  }
  parseRequest(body) {
    /* ... */
  }
  formatResponse(decision, context) {
    /* ... */
  }
  getToolName(body) {
    /* ... */
  }
  isQuestion(body) {
    /* ... */
  }
  parseNotification(body) {
    /* ... */
  }
}

Architecture

bin/farmer.js          CLI entry point (start/stop/status)
lib/server.js          Core HTTP + SSE server
lib/adapters/          Agent adapter interface + Claude Code adapter
lib/persistence.js     State persistence (atomic write, debounced)
lib/connect.js         One-step hook installation and settings.json management
lib/security.js        Token auth, CSRF, CSP, PID lock, audit log
public/index.html      Dashboard (inline JS, no build step)

Zero dependencies

Farmer has zero npm dependencies. SSE for real-time streaming, polling as fallback. Everything is Node built-ins.

Part of the grainulation ecosystem

Tool	Role
wheat	Research engine -- grow structured evidence
farmer	Permission dashboard -- approve AI actions in real time
barn	Shared tools -- templates, validators, sprint detection
mill	Format conversion -- export to PDF, CSV, slides, 24 formats
silo	Knowledge storage -- reusable claim libraries and packs
harvest	Analytics -- cross-sprint patterns and prediction scoring
orchard	Orchestration -- multi-sprint coordination and dependencies
grainulation	Unified CLI -- single entry point to the ecosystem

License

MIT