Compound Product
https://github.com/snarktank/compound-product- Category
- AI Agents
- Rank
- No. 1084Tools index
Previous survey · No. 1089 ·
- Pricing
- Open Source
- Type
- AGENT
- Builder
- snarktank
- GitHub
- 542 stars
- Date
About
Self-improving product system that reads reports, identifies priorities, and autonomously implements fixes via Claude Code.
What it does
Compound Product turns a markdown report into a bounded development workflow. It selects one actionable issue, drafts requirements, splits the work into verifiable tasks, runs a coding agent iteratively, applies configured checks, commits successful changes, and opens a pull request for human review.
Why it's ranked here
The workflow has unusually concrete boundaries for autonomous coding: one task per iteration, configurable checks before commits, a fixed iteration ceiling, persistent progress notes, and pull requests instead of direct merges. Its usefulness still depends heavily on report quality, agent judgment, and strong project checks.
What's good
It supports dry runs, multiple model providers, custom analysis commands, and either Amp or Claude Code for implementation. Fresh agent contexts inherit state through Git history, task status, progress notes, and maintained project guidance. Frontend tasks explicitly require browser verification, while failed quality checks block commits.
Tradeoffs
Autonomy requires bypassing normal agent permission prompts, granting repository writes, shell execution, network access, pushes, and pull request creation. Setup also requires several command-line tools, authenticated Git hosting, and a configured model provider. The supplied repository text includes no project test suite, so the documented safeguards are not independently demonstrated here.
How to use it well
Use it for teams that already produce useful operational reports, maintain reliable automated checks, and review every generated pull request. Start with dry runs, isolate execution, restrict credentials, and monitor early runs. It fits small bugs and focused product improvements. It does not generate the source reports or replace broader product judgment, security controls, or human code review.
Technical notes+
install.sh copies shell scripts into the target repository, creates compound.config.json when absent, and installs PRD and task skills for detected agents. scripts/analyze-report.sh selects Vercel AI Gateway, Anthropic, OpenAI, or OpenRouter credentials and emits priority JSON. scripts/auto-compound.sh finds the newest report, creates a branch and PRD, invokes task conversion, runs scripts/loop.sh, pushes, and opens a PR through gh. scripts/loop.sh repeatedly invokes Amp with --dangerously-allow-all or Claude Code with --dangerously-skip-permissions, stopping on a completion marker or the configured iteration limit. skills/prd/SKILL.md defines self-clarification, scoped requirements, verifiable acceptance criteria, and browser checks for UI work.
Observed
- License
- MIT
- Primary language
- Bash shell scripting
- Install surface
- Repository installer copies automation scripts and configuration into an existing project
- Interface
- Command-line workflow driven through shell scripts
- Coding agents
- Execution loop supports Amp CLI and Claude Code
- Model providers
- Report analysis supports Vercel AI Gateway, Anthropic, OpenAI, and OpenRouter
- Required tooling
- Requires jq and authenticated GitHub CLI; browser acceptance checks require agent-browser
- Scheduling
- Includes a macOS launchd scheduling example
Read from README.md, AGENTS.md, install.sh, config.example.json, scripts/loop.sh, scripts/AGENTS.md, scripts/CLAUDE.md, scripts/prompt.md, examples/sample-prd.md, scripts/auto-compound.sh, examples/sample-report.md, scripts/analyze-report.sh, examples/sample-tasks.json, examples/com.compound.plist.example, skills/prd/SKILL.md.
What it can do
Read and parse software reports
Software reports and documentation → Structured analysis of report contents
Identify system priorities and issues
Parsed report data and system metrics → Prioritized list of issues and improvements
Generate code fixes autonomously
Identified issues and system requirements → Working code implementations
Implement fixes directly into codebase
Generated code fixes and target system → Updated software system with applied fixes
Monitor system performance for improvements
System metrics and performance data → Performance analysis and improvement recommendations
Execute self-improvement cycles
Current system state and identified optimization opportunities → Enhanced system capabilities and performance
Tags
Tech Stack
Comments (0)
No comments yet
Editorially curated, with community endorsements as a secondary signal. Corrections welcome.