Clockwork

Product Introduction

1. Definition: Clockwork is a macOS-native desktop application that functions as a scheduling and execution platform for AI coding agents. It is a local agent orchestration tool that integrates directly with a user's calendar to manage, run, and audit automated AI-driven development tasks.
2. Core Value Proposition: Clockwork exists to solve the visibility and accountability problem of recurring AI agent work. It provides a centralized, auditable home for unattended AI coding tasks—like dependency updates, test sweeps, and documentation checks—that typically get lost in terminal history, cron jobs, or CI logs. Its core value is bringing human oversight, cost control, and repeatable execution to autonomous AI coding workflows.

Main Features

1. Calendar-Based Scheduling: Clockwork uses a real calendar interface for scheduling agent jobs, supporting one-off and recurring events with policies for missed runs. This integrates agent time with human time on a single visual grid, providing clear visibility into what automated tasks ran and what requires manual approval.
2. Sandboxed, Isolated Execution: Each agent run executes in a dedicated, OS-sandboxed git worktree. On macOS, it utilizes Seatbelt sandbox profiles to restrict file system access, isolating writes to the run's specific worktree and making credentials in ~/.ssh or ~/.aws unreadable from within the sandbox, enhancing local execution security.
3. Human-in-the-Loop (HITL) & Budget Guards: The platform incorporates mandatory approval gates for risky actions (like force-pushes), pausing execution until the user reviews. It also enforces hard budget bounds including USD cost caps, token/turn limits, and wall-clock timeouts at the supervisor level, preventing uncontrolled spending or infinite loops.
4. Auditable Reporting & Search: Every agent run generates a detailed, human-readable report including the Git branch created, a diffstat of changes, the actual dollar cost, and a full transcript. All reports and transcripts are full-text searchable, creating an auditable history of an AI agent's work and decisions.
5. Multi-Provider Engine Support (BYOK): Clockwork offers provider-agnostic agent execution, automatically detecting and integrating with multiple AI coding CLI engines such as Claude Code, Codex CLI, OpenCode, and Hermes Agent. Users Bring Your Own Keys (BYOK), and Clockwork never marks up API costs, keeping billing separate.

Problems Solved

1. Pain Point: Lack of centralized management and audit trails for AI coding agents. Developers struggle to track, repeat, and review work done by autonomous AI agents scattered across ad-hoc scripts, cron jobs, and chat sessions.
2. Target Audience: Senior Software Engineers, Engineering Managers, and DevOps/Platform Engineers who are early adopters of AI-assisted development and need to operationalize repetitive coding tasks. Specifically, developers managing multiple code repositories who require automated maintenance, testing, and dependency management.
3. Use Cases:
   • Scheduled Dependency Audits: Automatically running a "Dependency Surgeon" agent weekly to review, group, test, and create pull requests for library updates.
   • Flaky Test Triage: Scheduling a "Test Doctor" agent nightly to analyze test suite failures, quarantine noisy tests, and propose fixes.
   • Documentation Synchronization: Using a "Docs Drift Checker" agent to periodically find discrepancies between code and documentation and draft corrections.
   • Repository Health Monitoring: Configuring a "Repo Health Monitor" agent to deliver a morning digest of stale branches, failing CI jobs, and new security advisories.

Unique Advantages

1. Differentiation: Unlike generic AI agent platforms or simple cron wrappers, Clockwork is a productized solution built specifically for the unattended AI coding workflow. It differs by combining a native calendar UI, mandatory safety interlocks (approvals/budgets), and genuine local sandboxing into a single focused tool, whereas competitors often lack integrated scheduling, robust isolation, or transparent cost reporting.
2. Key Innovation: Its transparent and verifiable security model is a key innovation. The core security boundaries—the Seatbelt profile, credential deny list, and environment allowlist—are open-sourced under Apache-2.0 and accompanied by tests that prove their efficacy. This allows technical users to audit the security claims themselves, building trust in a proprietary application.

Frequently Asked Questions (FAQ)

1. Is Clockwork a cloud service or a local application? Clockwork is a local macOS desktop application. All AI agent scheduling, execution, and sandboxing happens directly on your Mac, not in the cloud. It uses your local Git repositories and your own API keys.
2. How does Clockwork handle security and prevent agents from accessing sensitive data? Clockwork uses per-run macOS Seatbelt sandbox profiles to enforce strict filesystem isolation. Each job runs in its own sandboxed git worktree, with writes locked to that scope. Crucially, it blocks agent access to directories containing credentials (like ~/.ssh). The sandboxing code is open-source for verification.
3. What AI models or providers does Clockwork support? Clockwork is provider-agnostic and supports multiple AI coding CLI engines, including Anthropic's Claude Code, OpenAI's Codex CLI, OpenCode Interpreter, and Hermes Agent. It automatically detects installed CLI tools and lets you choose the engine per task.
4. How are costs controlled when running unattended AI agents? Clockwork implements enforced budget guardrails at the application level. You set USD soft caps, turn limits, and timeouts per job. The Clockwork supervisor enforces these limits—not the AI model—and provides an honest report of the actual cost incurred for each run.
5. Can Clockwork run agents overnight or when my Mac is asleep? Clockwork requires your machine to be awake. It will keep a plugged-in Mac awake during scheduled run windows and alert you if a sleep state causes a missed job. For true 24/7 unattended operation, you need to run it on an always-on Mac (like a dedicated Mac mini).