Office Hours XIII: Ponytail, Ten Hermes Instances, a Web-MCP Verification Layer, and CodeRabbit

A GitHub repo installed into Claude Code / Codex-style environments that distils large code to the essential, executed lines before the agent works with it - 'it removes unnecessary code.' Reported ~122k installs and 'really good' over a few weeks of use. Harshit's second use for it: run Ponytail inside GitHub Actions on each push and you 'have your very own code reviewer tool', a free alternative to CodeRabbit for catching orphan or unfinished code.

Praful runs Hermes personas by project: website design (Open Design rather than Claude Design), app-building (Ponytail plus a referral/integration tool heard as Nango), a personal group, and a VPS-access group pairing Tailscale (visible remote) with computer-use (headless) - two mutually exclusive ways in. Claims: ~10 instances concurrently 'never ran out of the tokens'; ~20 client proposal sites in three hours; two Hermes instances on one VPS via Docker, dismissing 'VPS is restrictive' warnings - 'always a way in which you can get what you want to do.' Fable 5.1 is reserved for calculation-heavy research. Before any GitHub repo is installed, repo-safety skills (as-heard Skill Dart / sh skills; also Bumblebee, 'beads') vet it; Skill Router and Nvidia's Skillscape are name-checked. A 'Grokboard' (as-heard) session was planned for Sept 11.

'The web is treating the text file very separately and the video files are not' - agents and indexes ingest text and ignore video, though video now carries updates that supersede published papers. Cody's hackathon entry (Microsoft/Google/OpenAI-sponsored Web-MCP hackathon, built with a partner in ~3 of 7 days, ~28 hours of Codex 'emptying buckets everywhere') cross-references a video's claims against its linked source and flags which is current - a knowledge-integrity layer, with an interactive constellation UI. Roadmap: verification (done), ranking (next), then client/server architecture; open to community input because 'nothing is out there yet.' Harshit's feedback: pair it with a PostHog-style dashboard of Web-MCP usage wired to Hermes over ACP; Cody's counter - an app-drawer 'focus' UI that archives inactive projects as a token firewall.

Design sites so an agent can call intent directly - an MCP-style endpoint with pre-validated actions and the 'top 3 most likely actions' surfaced first - rather than parsing human UI, 'which is impossible for it.' Effect: 'latency would go down by a lot and the GEO... would increase automatically' - GEO as generative/agent engine optimization by analogy with SEO. Harshit frames it as a service AI/SEO agencies could sell but mostly lack tooling for.

Cody's Hermes agent controls his Android TV (Android required - Samsung and a couple of brands excluded) via a small dashboard on a Cloudflare Worker bridging HTTPS to the TV, with Wispr Flow on the phone for voice input. Basic commands work; the on-screen keyboard grids 'are all these walls' - a vision model to let the agent see the screen is next. Plan: extend to every same-Wi-Fi device (home assistant, CCTV) and split speech by region - Sarvam for Indian accents, Deepgram or Grok for Western. Google STT/Gemini did not interlace across his devices; Harshit reports smooth multilingual results (Hindi, Marathi, Gujarati, English, Bengali) from a recent Gemini-based test. Clarified: Hermes' fallback mechanism is a model router - to use one model for STT/TTS and another for tool calls you send discrete turns, not a mid-turn split.

To the chat question about tools that check a finished build for orphan or unfinished code, Harshit names CodeRabbit AI - 'I use that tool a lot' - citing its bus ads in San Francisco and Bangalore offices as traction signals. Free alternative: the Ponytail repo running in GitHub Actions as a homegrown reviewer pipeline. Aside: Sarah's screenshot-level bug report on his newsletter's email rendering was thanked as the fix that lifted his open rate.