Open Source Projects

A free Claude Code plugin that runs an approved, multi-task plan one task at a time, with a fresh reviewer checking every task before it counts as done.

Built after watching a single subagent hand back a plan it graded as finished when it wasn't. dispatch splits the work into a fresh implementer subagent per task, a fresh spec reviewer, and a fresh quality reviewer, none of which share context, then a gate that runs the plan's own verification commands and checks their exit codes rather than trusting anyone's report.

• One task at a time, in order, never two implementers running in parallel on the same files
• Fresh reviewer subagents for spec compliance and code quality, each reading the actual diff, not the implementer's account of it
• An exit-code gate that runs the plan's own Gates: commands after every task, before the checkbox is ticked
• A local ledger, one TSV row per task, so loop counts and escalation rates are visible over time instead of living only in chat history
• One approval per plan: a single design-pass approval, then the run proceeds unattended until it finishes or needs a human

Product page and eval results · Source on GitHub

A free, MIT-licensed menu bar app for macOS and a system tray app for Windows: limits, spend, subscriptions, every MCP server and agent running, and a scored audit of your AI setup.

The largest of the open source projects on this page. One codebase, built with Tauri (a Rust core and the system webview), reads what your AI tools already keep on your own machine and turns it into limits with forecasts, spend by client or work area, a subscriptions ledger, and a setup audit, without sending anything off the computer aside from two opt-in calls (a daily MCP Trust Index lookup and a GitHub update check).

• Two builds, one license: the personal menu bar / tray app, and a headless team half that reports a policy-checked seat inventory to a collector you host yourself
• No telemetry, no account: reads local logs and config files only; a live browser demo runs the real interface on fictional data
• MCP inventory across seven AI apps, limits for 23 providers, one-click package pinning, and an Agents view for what is running and what it cost
• Nine languages, translated throughout the app
• Standing on two other MIT projects: Pane by Jazii (the original Windows tray app) and OpenUsage by Robin Ebers (the macOS original)

Product page and live demo · Source on GitHub

105 AI agents, 20 playbooks, production-ready generators and quality gates.

A comprehensive framework for building, deploying, and managing AI agents at scale. Born from Peter's experience running 4 autonomous agents (Halperbot, Saarvis, MiniDoge, Nyx) that manage his VC fund, 10+ websites, and daily operations.

• 105 role-specific agents across 12 departments (engineering, research, ops, security, content, analytics, etc.)
• 20 operational playbooks — battle-tested workflows for agent deployment
• Quality gates — automated checks that prevent agents from shipping bad output
• Generator scripts — create new agents from templates with consistent structure
• Multi-provider LLM support — works with Gemini, Groq, Cerebras, Claude, and more

5-provider LLM failover library with automatic provider rotation.

When one LLM provider hits rate limits or goes down, llm-failover automatically routes to the next provider in the chain. Used in production to ensure 24/7 uptime across Peter's agent fleet.

• Provider chain: Groq → Gemini → Cerebras → SambaNova → Cloudflare Workers AI
• Catches all exceptions — not just rate limits, but timeouts, auth errors, and network failures
• Zero-config fallback — set your chain once, every call is resilient
• Battle-tested: handles 1,000+ daily LLM calls across 4 agents

38-check SEO scanner with zero dependencies, auto-fix, and snapshot diffs.

A Python CLI tool that crawls any website and audits it against 38 SEO checks. Includes an auto-fix mode that applies mechanical fixes (title truncation, meta description cleanup, JSON-LD syntax, missing viewport tags, and more).

• 38 checks covering meta tags, JSON-LD, OpenGraph, Twitter cards, accessibility, and technical SEO
• Auto-fix mode — applies safe, mechanical fixes with dry-run preview
• Snapshot diffs — compare audits over time to track regressions
• GEO audit — AI search visibility, citability scoring, brand mention scanning
• Used in production: weekly GHA workflow audits 9 sites every Sunday at 3am ET

Peter's open source philosophy: ship what you actually use. Every tool listed here runs in production, daily. No theoretical frameworks — just battle-hardened code that manages real businesses, real websites, and real money.