Pushary

Product Introduction

1. Definition: Pushary is a human-in-the-loop (HITL) platform and control panel for AI coding agents. It functions as a centralized decision-making layer that intercepts permission requests from AI agents, enabling remote approval via mobile devices and other channels.
2. Core Value Proposition: Pushary exists to eliminate workflow interruption and agent idle time by moving critical approval decisions from the developer's terminal to their pocket. Its primary value is enabling unattended AI agent execution while maintaining granular permission control and a cross-agent audit trail, thereby increasing developer productivity and agent safety.

Main Features

1. Universal Agent Integration via MCP: Pushary connects to AI agents like Claude Code, Codex, Cursor, and Hermes using the Model Context Protocol (MCP). A single npx command installs permission hooks and an MCP server, creating a unified control layer without requiring agent-specific API integrations. This allows it to work with any MCP-compatible agent.
2. Mobile-First, Keyless Approval Workflow: The core interaction is a QR-code-based pairing system. After a one-time scan, permission requests are pushed as native notifications to the iOS or Android app. Users can approve, deny, or reply directly from their lock screen, with the decision fed back to the paused agent in real-time. No API keys need to be manually copied or pasted.
3. Granular, Tool-Specific Permission Policies: Users can define permission presets (e.g., "Balanced," "Always Protected") that auto-approve low-risk actions (file reads, safe edits) while escalating high-risk commands (shell execution, network calls, deletions, deployments) to require manual approval. Policies include configurable wait times and automatic fallback actions (auto-deny).
4. Centralized Audit Trail & Fleet Dashboard: All decisions across all connected agents are logged into an immutable audit trail. A live Fleet Board provides an overview of active agents, pending requests, and recently completed actions, serving as a single pane of glass for AI agent governance and monitoring.
5. Multi-Channel Delivery & Webhook Integration: Beyond the mobile app, approval requests can be routed to Slack (with interactive buttons) or custom n8n workflows via webhooks. This allows teams to integrate agent approvals into existing collaboration and automation tools.

Problems Solved

1. Pain Point: Agent Blocking on Permission Prompts. The primary problem is productivity loss when an AI agent halts execution to ask for permission in the terminal, forcing the developer to be physically present. This breaks flow and prevents long-running or overnight tasks.
2. Target Audience: Software Developers and Engineers using AI coding assistants (Claude Code, Cursor), DevOps/Platform Engineers managing automated deployment or scripting agents, and Technical Team Leads requiring oversight and audit capabilities for AI-assisted workflows.
3. Use Cases: Unattended Refactoring/Code Generation: Starting a large code change and approving risky file writes from a phone while away. Overnight CI/CD Pipeline Management: Approving deployment steps for a nightly build process. Multi-Agent Coordination: Monitoring and controlling several agents working on different repositories or tasks from a single dashboard. Enforcing Security Policy: Ensuring no git force push or rm -rf command is executed without explicit manual review.

Unique Advantages

1. Differentiation: Unlike built-in "YOLO mode" or simple auto-approve scripts, Pushary provides context-aware, tool-specific policy control. It is also agent-agnostic, unlike vendor-specific solutions, creating a unified control panel for a heterogeneous AI agent environment. It prioritizes privacy by design, sending only decision metadata, not code transcripts, to its servers.
2. Key Innovation: The keyless, QR-code-based pairing mechanism for mobile devices drastically reduces setup friction compared to manual API key configuration. Its fail-closed architecture (where unanswered requests timeout to a defined safe state like "deny") and durable decision ledger provide robust safety and compliance guarantees for enterprise use.

Frequently Asked Questions (FAQ)

1. Does Pushary send my source code to its servers? No. Pushary is designed with a privacy-first architecture. It transmits only the metadata necessary for the decision: the question title, body, and the tool name (e.g., "Bash," "Filesystem"). Your actual source code and file contents never leave your local machine or CI environment.
2. What happens if I miss a notification on my phone? You configure a wait time and fallback action per policy. If a request is not manually approved or denied within the set timeframe (e.g., 30 seconds), Pushary will automatically execute the predefined rule, such as auto-denying the action or returning control to the terminal, preventing agents from being stuck indefinitely.
3. Can I use Pushary with AI agents I've built myself? Yes. Any AI agent that can be configured to use an MCP (Model Context Protocol) server can integrate with Pushary. For custom-built agents using frameworks like LangChain or the Vercel AI SDK, Pushary provides a decisions.ask() API that allows you to programmatically insert human decision points into any workflow.
4. How is Pushary different from getting Slack notifications for CI/CD approvals? While both deliver notifications, Pushary is agent-native and decision-blocking. The AI agent's execution thread pauses and waits synchronously for the decision from Pushary. It's not just a notification; it's a programmatic human-in-the-loop gate integrated directly into the agent's runtime, with a guaranteed response (manual or automated) required to proceed.
5. What is required to set up Pushary for an agent like Claude Code? Setup requires a single terminal command: npx @pushary/agent-hooks setup. This command configures the local MCP server, installs permission hooks, and links your environment to your Pushary account. You then scan the presented QR code with your phone to complete the pairing. No code changes to your projects are needed.