📱 they blocked OpenClaw. now what.
TL;DR
Anthropic cut off the old OpenClaw OAuth path, but Pete Steinberger immediately posted a CLI-based fix — the new command switches OpenClaw to use Claude Code’s local CLI backend instead of direct API auth, which Ray verified by inspecting the code and docs live.
If you keep using OpenClaw the “official” new way, you’ll start paying token-by-token — Ray showed his own Anthropic usage already ticking up by $6, and noted Max subscribers now get monthly credits plus optional prepaid top-ups like $50, $250, or $1,000 with discounts.
The CLI method appears to preserve session reuse and caching, not fire off a fresh expensive request every time — using DeepWiki on the repo, Ray traced how OpenClaw stores session IDs, resumes Claude sessions, and invalidates them only when prompts, profiles, or MCP config change.
DeepWiki was the sleeper tool of the stream — instead of manually reading the repo, Ray used it to ask “what does this command do?” and instantly got a code-level explanation of how OpenClaw swaps from
anthropic/...models toclaude-cliexecution.Ray used the second half of the stream to pressure-test agentic coding workflows, not just complain about Anthropic — he spun up Oh My Codex, CMUX/TMUX, and a spec-heavy “WatchClaw” project to show how parallel research, planning, and implementation might work in practice.
His bigger thesis is that specs are the real leverage in agentic engineering — pointing to Bellman and Seagrid’s viral Claude Code clean-room clone, Ray argued the magic wasn’t just model horsepower but having crisp behavior specs that multiple agents could execute in parallel.
The Breakdown
Anthropic flips the switch on OpenClaw
Ray opens with the news: Anthropic has blocked the old way OpenClaw users were piggybacking on Claude Code subscriptions through OAuth-style auth. He immediately grounds it in his own account, showing usage charges already appearing and explaining what this means if you were paying $100 or $200 a month and thought that covered everything.
Pete Steinberger drops a same-day escape hatch
The mood changes fast when Ray spots a fresh post from Pete Steinberger, the OpenClaw creator, recommending a different login flow: use Anthropic as provider, but set the method to CLI and make it default. Ray’s whole vibe here is, “if Pete says try it, we’re trying it,” and he starts testing it live, SSH’d into his machine, while chat helps catch a missing argument.
Repo archaeology with DeepWiki instead of guesswork
Rather than just trust the workaround, Ray clones the repo, opens Cursor, then gets tipped off by Tech Friend to use DeepWiki on the GitHub repo. That turns into the aha moment of the stream: DeepWiki explains that OpenClaw now swaps direct API calls for the local Claude binary, checks whether Claude CLI is authenticated, and rewrites model selection to claude-cli under the hood.
Why the CLI path might actually be the clean solution
Ray’s real concern isn’t just “does it work?” but “am I about to get wrecked on tokens?” He digs further and finds the important part: OpenClaw appears to keep session IDs, resume existing Claude sessions, and only invalidate them if prompts or configuration hashes change, which suggests Anthropic-style caching behavior should still hold.
Pricing, drama, and the one-week delay
He steps back to narrate the broader politics: Anthropic now wants OpenClaw users on direct paid usage unless they go through approved CLI usage, and prepaid credit bundles can get you discounts. Ray also recounts that Pete allegedly negotiated only a one-week delay, while hinting at the tension that Pete is now at OpenAI and Anthropic has copied some of his product ideas before.
The stream mutates into an agentic tooling jam session
Once the OpenClaw fire is mostly understood, Ray pivots into exploration mode: CMUX, TMUX, Hermes Agent, Gemma 4, and especially Oh My Codex. He ties this to his recent interview with Bellman and Seagrid, the Korean duo behind a clean-room Claude Code clone, and keeps coming back to the same point: these people aren’t just “using AI,” they’ve built orchestration loops that let agents research and implement in parallel.
WatchClaw: Ray’s own product-in-progress
Ray then opens up his own spec-heavy project, WatchClaw, a watch-first and menu-bar interface for monitoring and steering Claude Code, Codex, OpenClaw, and other agent sessions. The pitch is very Ray: tiny-screen constraints force clarity, and if you can control sessions from a wrist, glasses, or a $60 Raspberry Pi-style device, you get a cleaner interface than bloated desktop tooling.
Specs first, code later
The last chunk is basically a live demo of his workflow philosophy. He feeds WatchClaw’s long markdown spec into Oh My Codex using a “documentation only” deep interview flow, trying to get to a validated proving path before writing code; for Ray, the lesson from the viral Claude Code clone is that the leverage is in clear specs, acceptance criteria, and safe handoff loops like Ralph + Teams, not just smashing “generate” and hoping for the best.