Agent FM for Claude Code & Codex

Product Introduction

1. Definition: Agent FM is a local, open-source macOS desktop application that functions as a real-time ambient intelligence layer for AI-powered coding agents. It is technically categorized as a developer observability and workflow orchestration tool.
2. Core Value Proposition: It exists to solve developer context-switching and oversight fatigue by providing a passive, audio-driven feed of AI coding agent activity. Its core value is delivering ambient awareness of your Claude Code and Codex agents without requiring active terminal monitoring, enabling developers to stay in the loop on progress, blockers, and decisions in real time.

Main Features

1. Live Agent Narration (FM Stations): The app creates individual "radio stations" for each active AI coding agent (Claude Code or Codex). It uses a connected Large Language Model (Gemini or OpenAI API) to analyze raw session output—including tool calls, file reads, edits, and errors—and synthesizes concise, natural-language audio updates. This transforms technical logs into actionable, spoken status reports.
2. Global Mix & Project-Based View: Beyond single-agent stations, the Global Mix aggregates narration from all active agents across different projects into a single audio stream. The interface organizes agents by project (e.g., meridian-web, meridian-api), displaying timestamps, active files, and runtime status (Active, Thinking, Error, AwaitingUser), providing a high-level dashboard of all AI development activity.
3. Attention Routing & Blocker Surfacing: The system actively identifies and highlights moments where an agent requires human intervention. It surfaces specific attention requests, blockers, and approval questions (e.g., "needs your call on cookies vs server actions") in a dedicated panel. This ensures critical decision points are not missed in the ambient audio feed, acting as a prioritized notification system.
4. Local-First & BYOK (Bring Your Own Key) Architecture: Agent FM is an open-source Mac app with no cloud backend or telemetry. All session data processing is initiated from the user's machine. Users provide their own API key for a narration provider (Gemini or OpenAI), which is encrypted and stored locally. This means your code never leaves your machine to a third-party server, only the necessary context for narration is sent directly to your chosen provider.

Problems Solved

1. Pain Point: The cognitive overhead of constantly switching between an IDE/terminal and AI agent interfaces to check on progress, leading to disrupted focus and missed important status updates or blocking questions.
2. Target Audience: The primary user persona is a full-stack or frontend developer who extensively uses Claude Code and Codex for tasks like feature implementation, code refactoring, debugging, and writing documentation. A secondary persona is a technical lead or engineering manager who oversees multiple AI-assisted projects and needs a high-level overview of activity and bottlenecks.
3. Use Cases:
   • Parallel Agent Monitoring: Tracking multiple agents simultaneously working on different tasks (e.g., one refactoring a query layer, another implementing a new UI component) without manually toggling between sessions.
   • Context During Deep Work: Receiving passive, audio-based updates on agent progress while the developer is focused on a separate, complex coding task in their editor.
   • Blocking Issue Triage: Immediately being alerted when an agent encounters a permissions error, a design decision point, or a failed test, allowing for rapid intervention.
   • Project Onboarding & Audit: Gaining a quick, historical understanding of what AI agents have been working on in a codebase via the timestamped "Narration Logs" and project task list.

Unique Advantages

1. Differentiation: Unlike screen recording or log file scraping, Agent FM provides a curated, semantic layer over agent activity. Compared to simple desktop notifications, it offers a continuous, low-fidelity awareness channel (audio) that is less disruptive than visual pop-ups. It is not a management platform for the agents themselves but a dedicated observation tool.
2. Key Innovation: The product's core innovation is the conceptualization of AI agent output as a broadcast medium ("FM") and the application of real-time audio narration for developer observability. The technical implementation of passive session log parsing without proxying or slowing down the primary Claude Code/Codex sessions is a critical architectural advantage.

Frequently Asked Questions (FAQ)

1. Is Agent FM secure? Does my code get sent to Agent FM's servers? No, your code does not go to Agent FM servers. Agent FM is a local, open-source Mac app with no cloud backend. Some session context is sent directly to your chosen narration provider (Gemini or OpenAI) to generate audio, but this is done using your own API key. The app has no telemetry.
2. Will using Agent FM slow down my Claude Code or Codex sessions? No. Agent FM operates as a passive observer in a separate process, reading session output logs. It does not intercept, proxy, or wrap your agent's communications, so it introduces no latency to your primary coding workflows.
3. What is the cost of using Agent FM? The Agent FM application itself is free and open-source. The only cost is for the tokens consumed by the narration provider (Gemini or OpenAI) you connect, billed directly by that provider according to their pricing.
4. Which AI coding agents does Agent FM currently support? Agent FM currently supports Claude Code and Codex. Support for other agents, like Cursor, is indicated as being on the product roadmap.
5. Do I need to change how I use Claude Code or my editor? No. Agent FM is designed as a companion tool. You continue using Claude Code, Codex, and your editor exactly as you do now. Agent FM sits alongside them, providing ambient awareness without requiring workflow changes.