📰 Dev.to · CodeKing

My local AI assistant could capture screenshots just fine, but Feishu and Telegram still received no image. The real bug was an upstream channel capability gate

Keeping an AI coding session alive is not enough. The real improvement came when my agent started recalling successful workflows, rules, and facts instead of re

Three habits that make AI coding agents genuinely useful: give them a way to verify their own work, keep each task smaller than feels natural, and write down th

Chat-based AI assistants feel useful until they need you to stay online. I fixed that by letting scheduled tasks wake the assistant up in the background and rou

Parallel AI tasks looked impressive until two of them tried to drive the same desktop. I fixed it with resource-aware queuing instead of blind cancellation.

I could already send coding tasks to my assistant, but real workflows still broke at the browser and desktop layer. Here's the local desktop-control loop I adde

I built CliGate so coding agents can use runtime sessions, local skills, and desktop controls from one local assistant.

DeepSeek's lower API prices make model routing more practical for Claude Code, Codex, and ChatGPT-style coding workflows.

Treating web chat as a special page instead of a real channel split my agent state in two and caused silent delivery failures.

A practical model for AI coding agents that need to remember projects, tasks, executions, and runtime sessions separately.

A real bug from wiring Claude Code and Codex into DingTalk: the bot acknowledged work, but the runtime result disappeared unless I fixed delivery and fallback p

I got tired of switching between terminal agents, so I made Claude Code and Codex share one local chat workflow with direct runtime and assistant modes.

A short walkthrough of the CliGate dashboard flow that installs and configures Claude Code, Codex, Gemini CLI, and OpenClaw from one local panel.

I stopped letting my product assistant depend on whatever account or API key happened to win routing, and gave it an explicit model-source chain with fallbacks.

I stopped treating my README like a dumping ground and split my local AI gateway docs into a landing page, a docs hub, and an in-product manual.

I got tired of AI coding tools assuming model IDs and provider deployment names should match, so I put one local routing layer in the middle.

I wired Antigravity accounts into CliGate and learned the hard way that OAuth success, token refresh, project IDs, and chat source wiring are four separate prob

I wanted Claude Code to use Google Cloud billing instead of another Anthropic key, so I routed it through Vertex AI with one local gateway.

I got tired of three incompatible AI tool protocols, so I built a local Node.js proxy that translates requests, streams, and images.

How to route Claude Code to Claude Sonnet 4.6 and Opus 4.6 through Google's Antigravity using a Google login — no Anthropic API key, no manual config rewires.

The agent I actually needed wasn't another coder — it was a dispatcher. Here's the control-plane vs execution-plane split I landed on and why.

Adding DingTalk as the third channel to a local AI gateway forced me to confront whether my abstraction was real — or just two hard-coded integrations wearing a

CliGate is a local multi-protocol AI gateway that unifies Claude Code, Codex CLI, Gemini CLI, Telegram, Feishu, and DingTalk with account pools, app routing, st

Step-by-step guide to launching Codex and Claude Code sessions from a Telegram bot, with approval bridging and session continuity.