Skip to content

Pygg 0.1.0-dev.2 (developer preview)

Pre-release
Pre-release

Choose a tag to compare

@nambean nambean released this 19 Nov 03:30

Developer preview. Pygg runs your coding agents (Claude Code, Codex, OpenCode and any OpenAI- or Anthropic-format client) on cheaper models without losing the answers: a local gateway on Windows that routes each request to a provider you connected, falls back when one says no, and records what every request cost.

Download

  • Pygg_0.1.0_x64-setup.exe: the installer for Windows 10 and 11 (x64). It installs for the current user, with no administrator rights needed. Install it over 0.1.0-dev.1: your connections, keys and ledger live in your user profile and are kept.
  • Pygg_0.1.0_x64_en-US.msi: the same app as an MSI, for managed installs.
  • The command line: npm install -g @pygg/cli@dev (npm), for Windows, macOS and Linux.

The builds are not code-signed, so Windows SmartScreen warns on first run: choose More info → Run anyway. Check the download against SHA256SUMS.txt. Opening the app moves your store from schema v31 to v40; a store moved forward cannot be opened by dev.1 again.

New since 0.1.0-dev.1

  • Limits per session and per day. Each Pygg API key can have, besides its monthly limit, a limit per routing session and per UTC day. Once one is reached, the next request is refused with 402 and a message saying which limit and what to change; a request already running is never cut. Endpoint → API Keys → Limits…, or pygg gateway client --session-usd N --day-usd N.
  • A receipt for every launch. pygg run claude|codex|opencode prints, when the tool exits, what that launch cost, its cache share, the models that answered, fallbacks, and what changed in the Git repository it ran in. pygg launches --by-repo sums it per repository and week; pygg run --statusline claude shows the cost live in Claude Code's status line.
  • Claude Code's requests by kind: main, subagent, compaction and side requests, with how many subagents one prompt started (Usage → Token Mix, pygg usage).
  • Prompt cache: a session goes back to its first model after a fallback once no warm cache is lost; the cache marker now also reaches Anthropic's models on OpenRouter (99.7% of later prompts read from cache in a live check); each answer that missed the cache records its likely cause; the size of each session's tools and system prompt is metered. A longer cache lifetime and clearing old tool results are measured in shadow; nothing is changed in your requests.
  • Codex: pygg setup codex writes a model catalog, so Codex gets apply_patch and the real context size of the models Pygg lists.
  • Skills & Prompts: one library of skills in the Agent Skills format, installed into Claude Code's, Codex's and OpenCode's own folders, globally or in a project, with a count of what each agent sends with every request.
  • The command line: pygg chat and pygg ask with any connected model through the router; providers, models, combos, keys and the default model managed from the terminal.
  • Web sessions: pictures, videos, sounds and files a site makes come back with the reply.
  • Anthropic's own usage limit ("You have reached your specified API usage limits") holds that account until the date it gives, instead of being retried.

Not for unattended or financially critical use yet: limits rest on providers' prices and token estimates. Smart Tasks are experimental and on hold. What is built and how it was verified: Implementation Status.