Run Coding Agents From Your Phone With Telegram
Why Telegram works as a mobile control surface for coding agents, and how TelePi and TeleCodex turn voice notes, screenshots, and short prompts into real coding workflows.
Engineering Notes
Build notes & architecture breakdowns.
Deep dives into how Ora and htmlctl are designed, shipped, and iterated — from system prompts to deployment pipelines.
TelePi vs TeleCodex
A practical comparison of the two Telegram coding-agent bridges: context model, mobile UX, ASR, attachments, launch profiles, and when each one fits better.
TeleCodex: A Telegram Remote Control for OpenAI Codex
TeleCodex turns Codex into a mobile-first workflow: voice transcription, launch-profile switching, per-topic sessions, file exchange, and instant handback to the CLI from Telegram.
AI, the UI Anywhere
The broader essay behind TelePi and TeleCodex: why steering coding agents from your phone makes sense once most of the job becomes supervision instead of typing.
TelePi: A Telegram Remote Control for the Pi Coding Agent
After canceling Claude Code, I doubled down on the Pi coding harness and built a Telegram bridge — bi-directional session hand-off, cross-workspace switching, and model selection, all from a chat interface I was already using.
Why I Switched from Claude Code and Google AI Pro to GitHub Copilot Pro+
Claude Code's shared usage limits were a bad fit for my architecture-first workflow, Google AI Pro felt siloed, and GitHub Copilot Pro+ gave me a broader multi-model bench for planner, coder, and reviewer roles.
How a Newsletter Became htmlctl's Extension System
Restoring the newsletter forced more than a backend. It led to a reusable extension model with same-origin routing, compatibility validation, safer runtime boundaries, and static releases left intact.
We Don't Just Design for Humans Anymore
I tried to share an htmlctl update on social media. The preview was blank. That moment led to shipping automatic OG image generation, favicon support, robots.txt, and a sitemap generator — making discoverability effortless by default.
The Bar Has Been Raised: Consumer Expectations in the Age of AI Software
Ora is a good reminder of how fast the baseline is moving. If a solo developer can build a voice-first tool that understands speech, uses local or cloud models, remembers context, and acts across Mac apps, users will soon expect that kind of flexibility everywhere.
Dynamic Tool Discovery: How We Reduced Prompt Tokens by 60%
Ora registers 41 tools. Sending all their schemas to the LLM on every turn was expensive and noisy. Dynamic tool discovery introduced lazy loading via a client-side discovery index — shrinking the initial prompt tool block from ~1,200 tokens to ~450.
More posts coming soon · follow on GitHub