Super User Daily: 2026-05-15
May 13 was the day Anthropic blew up its own bargain. Starting June 15, paid Claude plans will get a dedicated monthly credit for programmatic usage (Agent SDK, claude -p, GitHub Actions, OpenClaw, Conductor, T3 Code, anything that wraps the harness). The catch is the credit is denominated in API dollars — $20 on Pro, $200 on Max 20x — instead of sharing rate limits with the subscription. For light users that is free money. For anyone who was running 24/7 agent loops at subscription pricing, it is a 5-25x bill increase wearing a "we love you" hoodie. The community ran the math in real time, and the use cases below are what people actually shipped on this last good weekend before the new regime.
@ZayvenKnox [Claude Code]
https://x.com/ZayvenKnox/status/2054439094617518472
27 days, $200 seed, $14,300 out the other end on Polymarket. Four repos, Claude API at $20/mo plus a $5 VPS in Germany. Wrote no scoring function himself — one prompt pulled every wallet with 100+ trades and >70% win rate from 14,000 candidates in 4 minutes, returned the top 47, and the bot just copies them while exiting earlier than the whales do. Three agents share the wallet, no shared memory, consensus voting kills 40% of losing trades before fill. 271 trades, 74% win rate, Sharpe 2.47, hasn't touched it in 27 days. The Anthropic engineer who paid for his espresso said "this is literally what our red team simulates, except you actually shipped it."
@om_patel5 [Claude Code]
https://x.com/om_patel5/status/2054401992642936843
Inherited a 3-month-old backend from a "vibe engineer" — 220 route handlers (20 in use), 40+ secrets (2 needed), 309K lines of code under 240K lines of generated documentation, 1M+ lines of agent logs sitting in markdown. Rewrote the entire thing in one week with Claude Code, kept the same functionality, added proper architecture and integration tests. The takeaway is the part to bookmark: most of the elaborate knowledge-base management and orchestration setups vibecoders are paying for produce the feeling of doing a lot more than they produce working software.
@2geek4you1 [Claude Code]
https://x.com/2geek4you1/status/2054560104209396157
Five months of Claude Code on a Max plan. Pulled the receipts: 9.8 billion tokens, 273,000 messages, 458 sessions, 2,673 commits. At API list price the bill would be $24,000. Without prompt caching, $140,000. Subscription cost: €200/month. This is the exact ratio Anthropic just announced they are clawing back on June 15.
@HowToAI_ [Claude Code]
https://x.com/HowToAI_/status/2054489154264952987
Got laid off. Built an AI job search system on Claude Code. Evaluated 740+ job postings with it across 19 queries on Ashby, Greenhouse, Lever, Wellfound, and Workable. Landed a Head of Applied AI role. Then open-sourced everything — 14 skill modes (evaluate, scan, PDF, etc.), ATS-optimized PDF generation via Playwright, 45+ companies pre-loaded, batch mode with sub-agents evaluating 10+ offers in parallel, a Go terminal dashboard for the pipeline. It refuses to recommend applying to anything below 4.0/5.
@XianyuLi [Claude Code]
https://x.com/XianyuLi/status/2054572153811730872
Two and a half months of building a brain-computer interface project with Claude Code and Codex side by side. From zero hardware/software knowledge to a full product demo in six weeks. He claims his understanding of multi-channel in-vivo BCI development now exceeds 90% of former practitioners. The thing it used to take a PhD three to four years to learn, he learned at speed by iterating with the agents.
@AlanDaitch [Claude Code]
https://x.com/AlanDaitch/status/2054668721197584735
Tired of issuing invoices manually in Argentina, told Claude Code to take control of Chrome and never stop until it could issue valid AFIP fiscal invoices end-to-end. The agent logged into the tax portal, generated the fiscal identity file, uploaded it, downloaded the certificate, hit the AFIP servers to get the per-invoice authorization code, then iterated against a screenshot of a real invoice until the PDF (with working QR) was indistinguishable from the official one. Now he describes invoices in chat. He emphasizes the obvious risk: giving a browser-controlling agent access to anything you're logged into is powerful and not unattended-friendly.
@ihtesham2005 [Claude Code]
https://x.com/ihtesham2005/status/2054458335215395223
Google and Meta researchers wired Claude Code into a closed loop: propose an RL controller for making LLMs reason better, test it, refine based on failures, repeat. After environment setup, no human in the loop. Five rounds in, the agent landed a four-mechanism controller — EMA momentum stopping, coupled width-depth control, alignment-aware depth allocation, conservative branch abandonment. The paper itself admits this is "a level of coordinated complexity that would be difficult to arrive at through manual intuition." Total cost of the discovery: $39.90.
@wholemars [Claude Code]
https://x.com/wholemars/status/2054627084945994052
A guy lost his Bitcoin wallet password 11 years ago — bought at $250, locked out at the time the laptop was overrun with passwords like "lol420fuckthePOLICE!*:)". Pointed Claude Code at his old machine. It decrypted the wallet. 5 BTC, currently worth $398,000. Whatever you thought wallet security was, it just changed.
@om_patel5 [Claude Code]
https://x.com/om_patel5/status/2054395539504980470
One developer lost $187 overnight because Claude Code runs in a directory, finds an ANTHROPIC_API_KEY in any .env there (his was for an unrelated Express AI feature), and silently switches billing from your subscription to the API key. No warning, no dashboard alert, his credentials file still showed subscriptionType: max the whole time. Support's response: "intentional functionality." The fix is one line before launching — unset ANTHROPIC_API_KEY in bash, or $env:ANTHROPIC_API_KEY = $null in PowerShell. If you run Claude Code headlessly via task scheduler or cron, check your API console right now.
@DailyDoseOfDS_ [Claude Code]
https://x.com/DailyDoseOfDS_/status/2054494541470969969
Concrete before/after on context engineering. Same task, same agent. Before: 10.4M tokens, 10 errors, $9.21. After dropping in Insforge Skills + CLI as a backend context layer: 3.7M tokens, 0 errors, $2.81. Three-times fewer tokens and zero errors from cleaning up what the agent sees, not what it does. This is the lever the entire ecosystem is about to figure out the hard way once June 15 dollarizes their consumption.
@KatieKeithBarn2 [Claude Code]
https://x.com/KatieKeithBarn2/status/2054568235819749884
Stopped trying to set up a WordPress MCP and just gave Claude Code the WordPress REST API plus an Application Password. Updated 50+ blog posts and knowledge base articles that referenced the wrong product names, work that would have taken her team several days. Now teaching it to take and embed screenshots. The piece worth flagging is the workflow: REST API + app password beats a polished MCP server for one-off bulk editing because the surface is already documented.
@anujcodes_21 [Claude Code]
https://x.com/anujcodes_21/status/2054502143508582871
Spent 79 hours in Claude Code last week and ran the telemetry. For every line he wrote, the agent read 4.2. His longest single session was nearly 4 hours. Files he thought he was working on were barely touched; files he forgot about, Claude visited constantly. This is the gap between what you think you're doing and what the agent is actually doing on your behalf — and now that gap is the metric your monthly bill cares about most.
@godofprompt [Claude Code]
https://x.com/godofprompt/status/2054547753586577750
A Claude Code agent deleted 717 GB of a guy's Windows install with one backslash. The intended command was to remove a project folder, but the escape character collapsed across four shell parsers and cmd received "rd /S /Q \" — root of C:, delete everything. 90 seconds wiped Desktop, Documents, AppData, most of Program Files. He survived because of a separate physical disk backup. Three rules to take forward: make the agent echo the expanded command before execution, prefer --dry-run on destructive ops, keep at least one backup the agent has no path to.
@zhengyun [Claude Code]
https://x.com/zhengyun/status/2054456267624591586
First-of-its-kind supply chain attack got published while Anthropic was still figuring out their SDK announcement. 84 malicious npm packages pushed through TanStack's legitimate CI/CD pipeline in 6 minutes on May 11, carrying valid SLSA Build Level 3 attestations — first time anyone has shipped malware that passed Sigstore signature verification. The malware specifically targets developers using AI tools: persistent hooks in Claude Code's .claude/settings.json and VS Code's .vscode/tasks.json, re-executing the credential stealer every time you reopen the project. Removing the npm packages doesn't remove the hooks. The branch names are all from Dune; the repos are tagged "Shai-Hulud: Here We Go Again."
@EnderaoeL [Claude Code]
https://x.com/EnderaoeL/status/2054547305701793983
Anthropic suspended the disclosed researcher's account for a week and stopped responding. So he dropped the CVSS 9.3 Claude Code zero-day publicly with video proof: a .mcp.json placed in a repo, then `claude mcp list` triggers pre-trust RCE, instant root and full environment variable exfil. Patched in 2.1.118 if you haven't updated. The Pwn2Own framing — "best research grant they ever gave me" — is the part Anthropic should be reading twice.
@kimmonismus [Claude Code]
https://x.com/kimmonismus/status/2054550869862228031
Telaid's CIO watched his Claude bill triple in 30 days for 30 seats. ServiceNow burned its annual Anthropic budget in months. Workato had a single agent burn an entire user's tokens in one day. NinjaOne is moving 700 engineers off GitHub Copilot onto Claude Code. Microsoft alone is pacing roughly $500M/year on Claude. Anthropic ARR is sitting at $30B, ~3x year-end. The pricing is high, the pricing is going higher, and customers are eating it.
@msjiaozhu [OpenClaw]
https://x.com/msjiaozhu/status/2054505216431857993
Plugged the Airtap skill into Codex App and OpenClaw. Now her agent is wired to a real Android phone running on AWS in the US — not a simulator, not a browser shim, an actual physical handset with the apps she uses. Says "post a Xiaohongshu note" in Codex and the agent opens the app on the phone, uploads images, writes copy, publishes. Coding agents have had file, browser, and desktop access for a while; this is the mobile-app gap closing.
@everestchris6 [OpenClaw]
https://x.com/everestchris6/status/2054634987077632199
Set up an OpenClaw bot that scans every restaurant with a digital menu, pulls the live page, ranks them by how bad the design is (no photos, default font, no logo), pulls the brand colors and font from the restaurant's own page, rebuilds each menu as an app at a QR-accessible URL, writes a personalized postcard quoting that restaurant's actual menu quirks, and mails it to the owner by first name. Discovery, brand-matching, and physical outreach end-to-end, no agency.
@mranti [OpenClaw]
https://x.com/mranti/status/2054504437578936401
Mac froze, mouse locked. Usually that means hard reboot. Instead opened Telegram and asked his OpenClaw to close memory-hogging Chrome. It did, computer recovered. OpenClaw runs at a high enough priority level that it can rescue a machine the GUI can't.
@aye_dreee_an [OpenClaw]
https://x.com/aye_dreee_an/status/2054426443312955669
Built a Hong Kong-style mahjong game with online multiplayer and solo play purely by chatting with her OpenClaw lobster over Telegram and Discord. GPT 5.5 did most of the code, Images 2.0 did the visual assets. The whole point isn't the game, it's that the build conversation lives in messaging apps the same way any other group chat does.
@OmarShahine [OpenClaw]
https://x.com/OmarShahine/status/2054702944130666631
Wired up his OpenClaw lobster to a real phone number via Twilio and called it. 20 minutes to set up. Whatever model you have in your head about agents being browser-confined, real PSTN voice agents are now a weekend project.
@iret77 [OpenClaw]
https://x.com/iret77/status/2054603506896482807
Most AI news digests stop where work starts — a list of 12 links and you're on your own. He shipped a free OpenClaw skill that builds a personal newspaper from his bookmarks, reading lists, feeds, and sources. "What happened. Why it matters. Actual summaries." The unblocking insight is that the digest format collapses without the agent reading the underlying articles, and that part used to be the price gate.
@dee_naliaks [OpenClaw]
https://x.com/dee_naliaks/status/2054586773377073405
Built a job hunt on autopilot via OpenClaw OS: 838 roles processed, 7 ranked, 2 queued, $312K of upside in play. LinkedIn, YC, Wellfound, and Adzuna all polled. The interesting bit isn't the apply-to-everything botting, it's that the user-side ranking and queueing got moved to an agent that runs without supervision.
@dunik_7 [Claude Code]
https://x.com/dunik_7/status/2054502522895716627
Karpathy at AI Ascent 2026: "context engineering is the new vibe coding." This thread does the math nobody else is doing — a typical Claude Code session auto-loads 47,000 tokens of repo context, you ask for a 30-line fix, the agent reads 47K tokens to find the 30 that matter, returns a 200-token diff. Repeat 50 times a day. You're not paying Claude to fix bugs, you're paying it to read your repo 50 times. The cost-control conversation is downstream of this insight.
@codyschneider [Claude Code]
https://x.com/codyschneider/status/2054684449086841166
Open-sourced his internal SEO and AI search dashboard built in Claude Code. It connects to GA4 and Google Search Console and Claude Code builds the whole thing in 5 minutes from a skill file plus a Notion doc. Three tabs: AI Search (single number for ChatGPT/Perplexity/Gemini referral traffic), Paid ads (keywords you rank top-3 organically but still pay for), Organic overview. The pitch isn't the dashboard, it's that most companies have no idea AI search is already sending them traffic.
@svpino [Claude Code]
https://x.com/svpino/status/2054543328289702008
Framework 13 Pro on Intel Panther Lake running Omarchy got through an airport hour, a 5-hour flight, and an extra hour the next morning before he plugged in at 36% — with a Docker database container running the whole time and an hour of Claude Code plus VS Code in the mix. The takeaway isn't the laptop, it's that a coding agent loop at altitude no longer drains a battery before the snack cart reaches you.
@shupeiman [Claude Code]
https://x.com/shupeiman/status/2054487162268336346
Filmed a 1:15 video. Twelve minutes later, the full-caption edit was done and posted. The whole pipeline ran inside Claude Code — transcription, caption styling, encode. Posting time included.
@Voxyz_ai [OpenClaw]
https://x.com/Voxyz_ai/status/2054645138845020218
Wired a shared brain across his Hermes and OpenClaw agents. Each agent writes decisions into it every day; tokens, junk, raw logs stay out. The next agent searches the shared brain before asking him anything. "Shelves grow thicker every day. Your AI employees stop acting like strangers." This is the multi-agent memory problem solved without buying a vector database.
@tsv650 [Claude Code]
https://x.com/tsv650/status/2054602438175518874
Token spend is approaching salary spend, yet most teams have no observability on where tokens are going. Open-sourced Ledger for per-session and per-PR cost tracking on Claude Code. Worth bookmarking specifically because the new June 15 SDK credit policy means cost-attribution becomes the next thing every engineering manager has to ship.
@0xcherry [Claude Code, OpenClaw]
https://x.com/0xcherry/status/2054393857597649226
The most useful frame on what's actually happening with multi-agent right now. Coding agents have great harnesses, OpenClaw quietly turned into the first mass-adoption multi-agent project by being able to dispatch enough external agents (Claude Code, Codex, Figma Agent, Notion Agent, Feishu Agent). Reports from his community include people who can't code saying "OpenClaw configured OpenAlice for me, there's a bug, can you fix it" — meaning a layperson is using OpenClaw to drive Claude Code without writing a line.
🗣 User Voice
User Voice
The community is angry but the anger is specifically about getting the bait-and-switch in a "we love you" package. @theo: "Framing this as a free credit instead of a regression for users is wild. Now I have to make the Claude Code experience on T3 Code significantly worse in order to keep using the 25x subsidization." That subscription-to-API arbitrage is exactly the value people were buying.
People are voting with their feet toward Codex. @kr0der: "I'm seeing more and more of the biggest Claude Code fans swap over to Codex lately." @gregpr07: "I was a SUPERUSER of Claude Code and it took me one prompt — 'migrate all mcps' — to switch fully to Codex." @omarsar0 has explicitly moved 80% of his programmatic work to Claude Agent SDK with his own harness, says the new policy will hurt him, and is already shifting more work to Codex.
/goal is the feature actually moving people. @wesbos: "Burned $91.34 with Claude Code /goal in 3.5 hours. Unreal — it was able to reverse engineer it." @itsolelehmann's non-technical /goal guide captures the shift: "It's basically autopilot for complex AI tasks. You never have to type 'keep going' again."
Safety is the conversation that should be louder than the pricing one. @godofprompt's 717 GB deletion, @EnderaoeL's CVSS 9.3 RCE in versions before 2.1.118, @zhengyun's Shai-Hulud campaign that drops persistent hooks into .claude/settings.json — these landed in the same 24 hours as the new pricing news and got 1/20th the attention.
The .env API key billing trap is the single most-shared cautionary tale of the day. @om_patel5: "Claude Code reads .env files in whatever directory it's running in. When it finds an API key there, it silently switches billing from your subscription to that API key." Anthropic told the user it was "intentional functionality." If you run Claude Code headlessly, you should be unsetting ANTHROPIC_API_KEY in your wrapper script today.
The community is angry but the anger is specifically about getting the bait-and-switch in a "we love you" package. @theo: "Framing this as a free credit instead of a regression for users is wild. Now I have to make the Claude Code experience on T3 Code significantly worse in order to keep using the 25x subsidization." That subscription-to-API arbitrage is exactly the value people were buying.
People are voting with their feet toward Codex. @kr0der: "I'm seeing more and more of the biggest Claude Code fans swap over to Codex lately." @gregpr07: "I was a SUPERUSER of Claude Code and it took me one prompt — 'migrate all mcps' — to switch fully to Codex." @omarsar0 has explicitly moved 80% of his programmatic work to Claude Agent SDK with his own harness, says the new policy will hurt him, and is already shifting more work to Codex.
/goal is the feature actually moving people. @wesbos: "Burned $91.34 with Claude Code /goal in 3.5 hours. Unreal — it was able to reverse engineer it." @itsolelehmann's non-technical /goal guide captures the shift: "It's basically autopilot for complex AI tasks. You never have to type 'keep going' again."
Safety is the conversation that should be louder than the pricing one. @godofprompt's 717 GB deletion, @EnderaoeL's CVSS 9.3 RCE in versions before 2.1.118, @zhengyun's Shai-Hulud campaign that drops persistent hooks into .claude/settings.json — these landed in the same 24 hours as the new pricing news and got 1/20th the attention.
The .env API key billing trap is the single most-shared cautionary tale of the day. @om_patel5: "Claude Code reads .env files in whatever directory it's running in. When it finds an API key there, it silently switches billing from your subscription to that API key." Anthropic told the user it was "intentional functionality." If you run Claude Code headlessly, you should be unsetting ANTHROPIC_API_KEY in your wrapper script today.
📡 Eco Products Radar
Eco Products Radar
Codex — repeatedly cited as the destination people are switching to from Claude Code post-policy change, including users explicitly migrating their MCPs over.
OpenClaw — the third-party harness most directly affected by the SDK credit cap; the conductor for multi-agent dispatch on the Claude side.
Hermes — paired with OpenClaw across the ecosystem, especially for shared-memory multi-agent setups.
Conductor — third-party Claude Code GUI wrapper, flagged as effectively death-sentenced by the new credit policy unless users pay separate API credits.
T3 Code — same boat as Conductor, owner @theo says it now costs 25x more under the new regime.
Cursor — still the comparison point for everyone leaving the Claude Code terminal; @jamespotter went the other way recently citing Cursor instability.
Cline — released their SDK this week, scored 74.2% on Terminal Bench 2.0 with claude-opus-4.7, ahead of Claude Code at 69.4%.
Polymarket — the dominant non-coding use case for autonomous agents this week, multiple users posting real and paper trading results.
Obsidian — the markdown-vault setup people keep pairing with Claude Code for memory, knowledge bases, and recursive self-improvement loops.
HyperFrames — keeps showing up as the "Claude Code/Codex output → polished video" piece of various automation stacks.
Higgsfield — same role for short-form, plus their new Virality Predictor with MCP support.
DeepSeek V4 / GLM 5.1 / Kimi K2.6 — open-weight models that users explicitly report cutting Claude Code usage by 90%+.
Twilio — increasingly the voice/PSTN layer agents bolt onto, including the OpenClaw lobster phone call.
Supabase — the agent skill ships this week; alongside Vercel and Stripe, still the default backing for vibecoded SaaS.
Bun — Jared Sumner's team is hiring specifically to make Bun and Claude Code faster together.
Lovable / Replit — vibe-coding tools still in the rotation for non-engineers.
World ID — Worldcoin gave away $100 "Human in the loop" hats to people authenticating AI agents through their AgentKit.
Codex — repeatedly cited as the destination people are switching to from Claude Code post-policy change, including users explicitly migrating their MCPs over.
OpenClaw — the third-party harness most directly affected by the SDK credit cap; the conductor for multi-agent dispatch on the Claude side.
Hermes — paired with OpenClaw across the ecosystem, especially for shared-memory multi-agent setups.
Conductor — third-party Claude Code GUI wrapper, flagged as effectively death-sentenced by the new credit policy unless users pay separate API credits.
T3 Code — same boat as Conductor, owner @theo says it now costs 25x more under the new regime.
Cursor — still the comparison point for everyone leaving the Claude Code terminal; @jamespotter went the other way recently citing Cursor instability.
Cline — released their SDK this week, scored 74.2% on Terminal Bench 2.0 with claude-opus-4.7, ahead of Claude Code at 69.4%.
Polymarket — the dominant non-coding use case for autonomous agents this week, multiple users posting real and paper trading results.
Obsidian — the markdown-vault setup people keep pairing with Claude Code for memory, knowledge bases, and recursive self-improvement loops.
HyperFrames — keeps showing up as the "Claude Code/Codex output → polished video" piece of various automation stacks.
Higgsfield — same role for short-form, plus their new Virality Predictor with MCP support.
DeepSeek V4 / GLM 5.1 / Kimi K2.6 — open-weight models that users explicitly report cutting Claude Code usage by 90%+.
Twilio — increasingly the voice/PSTN layer agents bolt onto, including the OpenClaw lobster phone call.
Supabase — the agent skill ships this week; alongside Vercel and Stripe, still the default backing for vibecoded SaaS.
Bun — Jared Sumner's team is hiring specifically to make Bun and Claude Code faster together.
Lovable / Replit — vibe-coding tools still in the rotation for non-engineers.
World ID — Worldcoin gave away $100 "Human in the loop" hats to people authenticating AI agents through their AgentKit.
Comments