Back to Projects

Claude Autonomous

Developer Tools1 weekSolo projectMarch 2026
DockerBashClaude CodelaunchdiptablesGoOrbStackGitHub Actions
One-command setup: Docker build, launchd install, git hooks
Budget enforcement: $100/day cap, 3-hour timeout, 500 max turns
Container firewall blocking private network access
Persistent task backlog with dependency tracking across runs
Successfully built VoxCoach end-to-end (47 tasks, 10 phases)

Claude Autonomous

Reusable Harness for Autonomous Development

Project Overview

Claude Autonomous is an open-source infrastructure harness for running Claude Code as a long-running autonomous agent inside an isolated Docker container on macOS. It schedules nightly development cycles via launchd, enforces budget and quality gates, and maintains persistent state between runs.

The Idea: Coding assistants are powerful interactively — but can they sustain long-term software projects autonomously with proper safety guardrails, budget control, and quality enforcement?

What We Built: A containerized environment where Claude Code runs on a schedule: picking tasks from a backlog, implementing with tests, committing only when quality gates pass, and documenting decisions for the next run.

Technical Architecture

Nightly Execution Cycle

  1. launchd triggers claude-nightly.sh at 2 AM daily
  2. Script refreshes OAuth credentials, injects operator instructions
  3. Docker container starts with workspace mounted
  4. Claude Code reads CLAUDE.md for context and tasks.json for backlog
  5. Agent picks next pending task, implements with tests
  6. Commits only if make check passes (fmt + vet + lint + test)
  7. Updates task status, logs output in structured JSON
  8. Budget enforcement: $100/run cap, 3-hour timeout, 500 max turns

Security & Isolation

  • Container Firewall — Blocks access to private networks, allows only HTTP/HTTPS/SSH
  • Permission Allowlist — Only approved bash commands (gofmt, go build, make check)
  • No Secrets in Container — Credentials injected at runtime, never persisted
  • Network Isolation — init-firewall.sh configures iptables rules

Persistent State

  • Task Backlog — tasks.json with dependencies and status tracking across runs
  • Agent Learnings — AGENTS.md documents patterns discovered across runs
  • Cost Tracking — JSONL logs for spend monitoring
  • Git History — All work auditable through conventional commits

Technology Stack

  • Container: Docker (Debian-based, Go 1.24, Node.js, Claude Code CLI)
  • Scheduling: macOS launchd (plist daemon)
  • Orchestration: Bash scripts with budget enforcement
  • Security: iptables firewall rules, permission allowlists
  • Quality Gates: golangci-lint v2, go test -race, gofmt, go vet
  • Git Hooks: Pre-commit (gofmt, gitleaks), commit-msg (conventional), pre-push (make check)

Key Features

  • One-Command Setup — ./setup.sh builds Docker image, installs launchd job, configures hooks
  • Budget Controls — $100/day spend limit with automatic shutdown
  • Quality Gates — Code only committed when all checks pass
  • Auditable — Every decision traceable through git history and structured logs
  • Reusable — Generic harness; swap the workspace for any project
  • Proven — Successfully built VoxCoach across 12+ nightly runs (47 tasks, 10 phases)

Links:

View Code Repository
Back to All Projects