The new X algorithm release has a blunt lesson for builders: The platform is not looking for “content.” It is looking for conversations that are hard to fake. A like is cheap. A repost is cheap. A thoughtful reply is expensive. If you build in public, stop posting announcements. Post the tiny decisions, failed attempts, weird benchmarks, and exact patches that make another builder say: “wait, how did you do that?”

Does anyone want to try it out?

**What it is:** buddy is a desktop app for people who run several AI agent CLIs in parallel (Claude Code, Codex, Gemini CLI, opencode). Each agent runs in its own PTY pane — split layouts, session history with one-click `--resume`, isolated login profiles per account (separate `CLAUDE_CONFIG_DIR` / `CODEX_HOME`), and a "Pulse" overview showing which agent is waiting for you. There's also a built-in chat that drives the CLIs headless (`claude -p --output-format stream-json` / `codex exec --json`), so it rides your existing subscriptions — no API keys. **Built with Claude Code, for Claude Code.** Full transparency: Claude wrote the large majority of this codebase — the Rust backend (Tauri 2, portable-pty, tokio, rusqlite) and the React/xterm.js frontend — across multi-day sessions, with me directing, testing on my machine, and feeding back screenshots when the UI broke. **Where Claude actually earned it** (the parts I couldn't have done quickly myself): - Diagnosed why sessions never showed "exited" on Windows: ConPTY doesn't EOF the reader when a child exits naturally → it added a dedicated waiter thread per session blocking on `child.wait()`. - Hit Rust's CVE-2024-24576 (BatBadBut) hardening in the wild: npm installs the CLIs as `.cmd` shims, and Rust refuses to spawn them with `"` in any argument — inline JSON flags silently can't work. It rewrote the flag to pass a settings file instead. - Fixed the classic "app works in terminal, broken from the dock" on macOS/Linux by resolving the login shell's PATH at startup (`$SHELL -ilc env`, with timeout fallbacks). - Found a UI bug from a screenshot I pasted: an invisible layout container was eating every click on the empty-state buttons. It traced hit-testing through the component tree and fixed it with one `pointer-events` line. The workflow that worked best: keep a persistent memory file of decisions, make Claude verify every pass (`cargo test` `clippy -D warnings` `tsc` vitest — currently ~45 Rust tests, 148 frontend tests), and paste real screenshots instead of describing bugs. **Free & open source:** github.com/ryan-mt/buddy — build it with `npm run tauri build`. v0.1, daily-driven on Windows; macOS/Linux hardened recently but less battle-tested. Feedback welcome.

done version 0.0.4 github.com/ryan-mt/tomte/rel…

tomte 0.0.3 is out. The agent now seals proof of its work onto the commit itself. tomte prove collects real exit codes from your tests, tomte seal writes them into a git note, seal verify gates CI from any clone. The model never grades its own homework. github.com/ryan-mt/tomte #Claude #tomte #BuildingInPublic

AGI man

I've been building tomte, a terminal coding agent written in Rust, and 0.0.2 is out. Sharing it for honest feedback. It's one binary you point at a repo. It reads, writes, runs, searches, and reasons through real work - streaming, with a 27-tool belt and a TUI that stays out of the way. You bring the model: sign in with a ChatGPT or Claude subscription (OAuth) or drop in an API key, and switch models mid-session. A few things that make it its own thing rather than yet-another-wrapper: Multi-model by design. OpenAI and Anthropic, OAuth or API key, in the same binary. /model swaps mid-session. It remembers why, across models. record_decision logs the reasoning behind a non-obvious change to a decision trail that's re-injected every session - so a later session, or a different model, inherits why it was done that way. Read it back with tomte why <file:line> / tomte blame <file>. Shell runs inside an OS sandbox. Default workspace-write with network off - Landlock seccomp on Linux, sandbox-exec on macOS - so a prompt-injected curl … | sh or rm -rf ~ can't reach the network or escape the workspace. A separate layer flags destructive commands and refuses them until you override. (Windows confinement is still best-effort - the built-in doctor tells you so; this release landed a big hardening pass on the command classifier.) Code intelligence, zero setup. A built-in lsp tool gives symbols / go-to-def / references / hover for Rust, TS/JS, Python and Go - no language server to install. Git worktrees on demand, live cost/usage/context readouts, project memory (CLAUDE.md/AGENTS.md), MCP servers, skills… and yes, it hatches a pixel companion in the corner. It's genuinely early (0.0.x), but it's real and I use it daily. MIT, prebuilt binaries for Linux/macOS/Windows. Repo releases: github.com/ryan-mt/tomte/rel… I'm the author - feedback very welcome, especially on the sandbox/permission model and the decision-trail idea. Happy to answer anything.

bao 🥟 a Claude Code skill that reviews code like a senior who's been burned before. reads the whole file, not just the diff. verifies every finding against the real code. if your change is clean, it just says so no made-up bugs. #ClaudeCode #buildinpublic github.com/ryan-mt/bao-revie…

🦀 opencli beta4 is out - a Rust coding-agent CLI, built as a drop-in replacement for Claude Code. Powered by the OpenAI Responses API (GPT-5.x). Sign in with ChatGPT (OAuth) or your API key, full tool-calling, streaming TUI. Linux/macOS/Windows · MIT · pre-release #RustLang #buildinpublic #AI github.com/ryan-mt/opencli

Why does Claude always add these ID models to my projects? It seems like there’s no filtering involved in them at all. @AnthropicAI @trq212 @bcherny

🦀 opencli beta4 is out - a Rust coding-agent CLI, built as a drop-in replacement for Claude Code. Powered by the OpenAI Responses API (GPT-5.x). Sign in with ChatGPT (OAuth) or your API key, full tool-calling, streaming TUI. Linux/macOS/Windows · MIT · pre-release #RustLang #buildinpublic #AI github.com/ryan-mt/opencli

Made a cute desktop app to watch your Claude Code usage 🐾 Live 5h / weekly limit gauges, burn-rate forecast, cost & token breakdowns, and an activity heatmap. Reads your real local data — the token stays on your machine. Electron React · Linux/macOS/Windows #ClaudeCode #Electron #AItools github.com/ryan-mt/claude-us…

Shipped OpenCLI v0.0.1-beta.1 🦀 A coding-agent CLI in Rust — built as a drop-in replacement for Claude Code, powered by OpenAI GPT-5.x. Sign in with ChatGPT or an API key. Prebuilt binaries for Linux, macOS & Windows. github.com/ryan-mt/opencli #Rust #AIcoding #OpenSource