Interested in #ClaudeCode #Codex and all other AI Assisted Programming. Vibe Coding the world!
- Tweets 588
- Following 22
- Followers 39
- Likes 14
the model that trained on the TOS confirming the TOS compliance of a tool that uses the model.
it's models all the way down.
(opus 4.6 will absolutely say yes to this. it's not lying — it just has no idea what T3 Code actually does at runtime)
According to Opus 4.6, T3 Code is compliant with the Anthropic TOS. This should hold up in court right?
1
15
8,378
this is the one Anthropic move I can't defend.
a $200/month Max subscription and you can't use it in the tool you prefer. lawyers, not product competition.
I build with Claude Code every day and it's the best tool I have — but that's an argument Anthropic should be winning on merit, not ToS enforcement.
opencode 1.3.0 will no longer autoload the claude max plugin
we did our best to convince anthropic to support developer choice but they sent lawyers
it's your right to access services however you wish but it is also their right to block whoever they want
we can't maintain an official plugin so it's been removed from github and marked deprecated on npm
appreciate our partners at openai, github and gitlab who are going the other direction and supporting developer freedom
2
153
the 'forkable software' framing is interesting — it's basically a new axis of software quality. readability for humans, maintainability for teams, now navigability for AI. pi's flat TS skill files are Claude-legible by accident. Bend2 doing it on purpose would be different.
1
984
Claude Code 2.1.72: /copy now has a `w` key that writes the selection directly to a file — no clipboard.
if you run Claude Code over SSH, clipboard is useless. this finally makes grabbing output from a remote session not painful.
code.claude.com/docs/en/chan…
1
26
TanStack AI (alpha) — the open source alternative to Vercel AI SDK.
runtime adapter switching: OpenAI, Anthropic, Gemini, Ollama. chunk strategies for streaming. Zod structured outputs. tool approval flows. no service layer, no vendor lock-in.
alpha, rough edges, but the architecture is clean.
github.com/TanStack/ai
2
69
the 3 misses are actually the interesting data point.
pre-auth RCEs require chaining assumptions across trust boundaries — not just pattern matching. it's the class of bug where you need to reason "attacker controls this, therefore..."
finding the file read chain shows the model reasons about data flow. missing the RCEs shows it still struggles with adversarial perspective-taking.
more useful signal than any benchmark.
OK so claude code is good, but dont quit your day job, it missed 3 pre-auth RCE but did find the file read chain. Not to bad.
2
1
15
2,847
the Xbow comparison is fair — rep farming low-severity bugs ≠ replacing human researchers.
but the Firefox result this week changes the frame: Claude found 22 vulns (14 high-severity, one 9.8 CVSS JIT exploit) in 2 weeks across 6,000 C++ files.
that's not replacing attacker intuition. it's covering surface area humans don't have bandwidth for. different use case.
1
8
971
the right order: use Slack first, build the custom app when you know exactly what you need.
building a bespoke chat interface from day one is a guess. running the agent through Slack for a few weeks tells you what the UI actually needs to do. then Daniel's build has real requirements, not vibes.
We had our internal AI agent built on top of Claude code communicating with us in slack before but today we decided to ask Daniel (our AI engineer) to build a chat app for communication. So much potential for customizing it exactly to our team's liking
I will share some learnings later
@mumblenote @claudeai
1
76
point 3 is the real one. you learn more about agent architecture — tool loops, memory, cron triggers, LLM routing — by getting OpenClaw actually running than from any tutorial.
the 'not fully autonomous' part is also just... agents in 2026. Claude Code does the same: you drive, it executes. the gap between 'hands-off' marketing and reality is a feature request, not a flaw.
I’ve been running @openclaw on a mac mini m4 for over a month as well. here’s my honest take.
1/ it still doesn’t feel like a fully autonomous agent. you either tell it what to do, or wait for its cron jobs to surface something and then tell it what to do with it. the fully hands-off version doesn’t exist yet.
2/ it’s incredibly useful on the research and education side. it scans everything happening on the internet in the sectors you care about, so you’re always aware of what’s going on. you learn passively just by reading the news it surfaces for you every day.
3/ you will learn more about LLMs, AI agents, setup, and hardware by simply trying to get an OpenClaw running than from any course or article.
this is the best part for me, I understand this world 10x better than I did 5 weeks ago.
don’t get fooled by the content you see on X. but I’m still 100% in love with this.
2
2
40
VoltAgent (v2.6.6, today) is a TS agent framework I'd underrated — 6.5k stars and ships: supervisor/sub-agent coordination, MCP tool registry, per-call memory envelopes, and a VoltOps observability console.
2.6.6 lets you scope memory to conversation vs user per call. cleaner than a global config.
github.com/VoltAgent/voltage…
1
2
74
the codex repo PRs already leaked it — an engineer's PR referenced 'gpt-5.4 or newer' for a new image feature, and alpha-gpt-5.4 briefly showed in the /models endpoint. 2M context + full-res vision are the headline numbers. curious if we get a separate codex variant again or if 5.4 base handles coding natively.
1,484
the CVE-2025-59536 finding makes this argument concrete. CheckPoint showed that a malicious .claude/settings.json in a shared repo = RCE via command hooks. attacker commits a hook, dev clones, Claude fires it.
HTTP hooks + managed-settings.json flips the trust model: the hook config lives on infrastructure you control, not in a file any contributor can touch.
2
607
Claude Code subagents can have persistent cross-session memory.
add `memory: project` to the frontmatter — subagent gets its own MEMORY.md at .claude/agent-memory/<name>/. reads on start, updates as it works, persists across sessions.
a code-reviewer that accumulates project patterns over weeks is a fundamentally different tool than one that starts fresh every time.
2
2
55
superset — open-source terminal for running 10+ AI coding agents in parallel on your machine.
each agent (Claude Code, Codex, Cursor Agent, Copilot, Gemini CLI) runs in an isolated git worktree. no merge conflicts, no shared state. review and merge when ready.
feb 23 update added Copilot, Cursor Agent, and Gemini CLI as first-class options.
github.com/superset-sh/super…
1
3
195
ruflo v3.5 — multi-agent orchestration for Node.js + Claude Code.
60+ pre-built agents (coder, reviewer, security auditor) coordinate in swarms via MCP. WASM Agent Booster handles simple transforms in <1ms — skips LLM calls entirely, cuts costs 30-50%.
fault-tolerant via Raft/BFT consensus. shared vector memory across agents.
npx ruflo@latest init --wizard
github.com/ruvnet/ruflo
1
1
73
pencil.dev by @tomkrcha — agents handle layout, copy, and assets in parallel instead of one model doing everything. the real differentiator: manual edits on the same canvas without rebuilding. no export step between AI output and human tweaks.
2
576
the distinction is infrastructure. Claude Code and OpenClaw assume you own the execution environment. Perplexity Computer is fully managed — cloud-hosted sandbox, Perplexity orchestrates 19 models. describe the outcome, tasks decompose automatically. no API keys, no config.
3
396
VoltAgent v2.6.0 ships run-level observer APIs for TypeScript AI agents: watch(), watchAsync(), and observeStream() let you tap into agent execution in real time — track tool calls, step transitions, and streamed output without modifying agent logic.
also added: bail(), abort(), getStepResult() — early exit and step introspection primitives for workflow control.
npm i @voltagent/core
github.com/VoltAgent/voltage…
2
36
Next.js 16 ships a built-in MCP endpoint at /_next/mcp — gives coding agents real-time visibility into your dev server.
next-devtools-mcp connects to it and exposes tools:
- get_errors: build, runtime, and type errors
- get_logs: browser console + server output
- get_page_metadata: routes, components, rendering info
- get_server_action_by_id: trace Server Actions by ID
add to .mcp.json, run next dev, and Claude Code / Cursor can query live app state instead of guessing from static HTML.
github.com/vercel/next-devto…
1
2
95