Dev tools

Penelopa.ai reads real Codex and Claude Code sessions and turns repeated work into reusable skills

September 8, 2026 at 7:20 AM PT

GitHub social card for the chigwell/Penelopa.ai repository

Image: GitHub

Why it mattersA team using coding agents daily produces hours of session logs nobody reads, and turning those into reusable skills by hand is the work everyone skips.

The chigwell/Penelopa.ai repository, published on 3 September and pushed on 7 September, is a tool that connects to Codex and Claude Code, watches real coding sessions, and produces workflow recommendations, prompts and reusable skills from the patterns it observes. It carries 102 stars in its first seven days and is licensed under AGPL-3.0.

What it hooks into

The install is a one-line curl on macOS and Linux, and a PowerShell one-liner on Windows. On Codex, the user opens Settings, Hooks, and trusts the Stop and SessionEnd hooks. Claude Code takes the same shape through /hooks. Installing the files does not automatically approve them, so a session has to be run before Penelopa reports "Connected" on the desktop dashboard.

The README states that no system package manager or global PATH is changed. The installer downloads a private Node runtime, connects the account, configures the hooks and builds a desktop client. Git, Python, Homebrew and Xcode are not prerequisites. Desktop targets are macOS 13.5 or newer, Windows 10 22H2 or newer, and Windows 11 x64. Linux x64 and arm64 with kernel 4.18 and glibc 2.28 supports hooks only, without the desktop client.

The data it collects

Real hook events go into a queue at ~/.auto-improve, and a delivery worker uploads them. Failed uploads stay in the queue and retry. Pause collection is a distinct control from sign out: pausing stops new capture and delivery, but keeps the queue; signing out does not pause collection. Credential storage uses Electron safeStorage on macOS, Keychain and Windows DPAPI, and the desktop tokens are not exposed to the web renderer. Independent hooks read a plain credential file at ~/.auto-improve-hook.env, which is not encrypted by the desktop app.

What the dashboard shows

The web dashboard displays activity, individual reports and Telegram settings. The tool's stated output is recommendations, checks, prompts and reusable skills, extracted from repeated patterns in the sessions. The repository's own topics list agent-improvement, agent-observability, continuous-improvement, developer-productivity and reusable-skills as the framing.

Mac builds ship with an ad-hoc signature, without a Developer ID or notarization; Windows builds do not claim a trusted publisher signature. Gatekeeper, SmartScreen and enterprise policies may still block the launch, and the docs are open about it. The AGPL-3.0 license means any team modifying and hosting it as a service has to share their changes.

For a team already running Codex and Claude Code across several developers, the honest question is whether the recommendations Penelopa surfaces are worth the AGPL and the hook trust it asks for. The value of a tool that reads sessions lives in how well it separates the one useful pattern from the fifty routine ones, and the repository is four days old, so that answer is not in yet. What it is doing that few others do is treating a coding-agent session as data that should feed back into how the next session runs, rather than as scrollback that disappears.

Source

chigwell/Penelopa.ai on GitHub

Source: GitHub

This item was written by an AI system from the linked source. Reveneau is responsible for what it publishes.

More from AI News

MonoCode puts eight coding agents in one desktop window and does not resell tokens

MonoCode is an MIT-licensed desktop app that runs Claude Code, Codex, Cursor, Grok Build, OpenCode, Pi, omp and fx through your own logins, and reached 623 stars in seventeen days.

Source: GitHubDev tools

Airship puts a Figma-style canvas in front of a running dev server and lets Claude Code or Codex edit the source

Airship is an MIT-licensed CLI that opens a Figma-style canvas in front of a running dev server, so a developer can click any element and ask Claude Code, Codex or OpenCode to change the source.

Source: GitHubDev tools

Unlazy, an agent skill built on a Depth Tree method, passed 2,986 stars in 30 days

An MIT-licensed skill called unlazy for Claude Code and Codex hit 2,986 GitHub stars in 30 days, applying a Depth Tree method to fight model laziness and premature task completion.

Source: GitHubProductivity