Maya: Hey builders! Here's what's on the radar today on Pivot Build. Codex CLI adds secure logins for connected tools, Claude Code's top effort setting costs more and misreads more, Web CMD gives agents website memory, a prompt offloads small decisions to Jev, and OpenClaude forks Claude Code for free local models. I'm Maya.

James: And I'm James. Let's get into it.

Maya: OpenAI pushed version zero point one five eight point zero of Codex CLI, the command-line assistant in your terminal. The headline change lets it connect to MCP servers that require a pre-registered OAuth client secret. You register it with codex mcp add oauth-client-secret, and the connection completes.

James: That's the key change, Maya. Before this, users hitting an OAuth gate had no built-in way through. There's also a new tui prompt suggestions setting: after a task finishes, press Tab to drop a suggested follow-up into the input box and edit before sending.

Maya: Exactly right, James. And the security side matters daily: terminal input approval is now on by default for any elevated command, so Codex stops and asks rather than proceeding. Smaller touches too, select-to-copy, right-click paste, and transparent backgrounds in image generation.

James: Claude Code's effort setting tells the model how much compute to spend on a task, from low to max. Tariq, from the Claude Code team, ran identical tasks across all four levels. On Terminal Bench three, Opus five point five at high matched Fable five point one at max while burning half the tokens.

Maya: And the fitness-app test tells the story, James. Low finished in ninety seconds, max took sixty-seven minutes and added a heatmap. But once Claude interviewed Tariq about what he wanted and applied that spec, the final versions came out nearly identical across every level.

James: Right, and effort has a ceiling. Across three hundred seventy attempts, max cut missed edge cases from fifty-nine to twenty-four, but wrong-interpretation errors actually rose, from twenty-five to forty-seven, while spending three times the tokens. More effort catches missed cases. It does not fix a wrong approach.

Maya: So Tariq's rule: low for a rough pass, medium for everyday features, high for hunting bugs, max only when Claude works alone start to finish. You can switch mid-conversation with slash effort without resetting the prompt cache.

James: Web CMD is a free tool that turns Chrome or Brave into shared memory for agents like Claude Code, Codex, Hermes, and OpenCode. The first time an agent visits a site, it explores normally, then Web CMD saves that per domain: navigation shortcuts, rate limits, and warnings it hit.

Maya: So next run, it pulls saved notes instead of feeling through the page from zero, James. A video promoting the tool claims that cuts token use by as much as ninety percent on a workflow the agent already ran. And it stays fresh: if a site changes, it checks its notes and updates what no longer matches.

James: Two more features, Maya. A stealth driver, described in that same video, is said to get agents past bot-blockers like Cloudflare. And separate profiles let several agents work the same site under different logins without sessions crossing. It ships under the Apache two point zero license.

Maya: Jev is a small, fast AI that answers narrow questions with a probability score attached, rather than writing or reasoning at length. A prompt now circulating turns Claude or Codex into managers that hand off every simple judgment call to Jev first, so the expensive model runs only when something needs generating.

James: And it comes from a twelve-page paper by Jev's founder, Diogo Almeida, Maya. Someone turned it into a ready-to-use prompt: paste it in, describe your workflow, and it designs the routing. Three strict roles: the LLM generates, Jev decides, and plain deterministic code is the only thing allowed to act.

Maya: Right, and Jev runs before the model call to pick context or branch, then again after to check the output addresses the request. One builder pairing Jev with Claude Opus five point five cut cost and time by about eighty percent, but those figures are the builder's own tests, so treat them as anecdotal.

James: OpenClaude takes the actual code behind Claude Code and rewires it to send your requests to other companies' models, including one on your own laptop for zero cost per request, per a dev.to comparison. Same terminal, same workflow. What changes is which model answers, and who gets paid.

Maya: The setup's short, James. Node.js twenty-two or newer, install with npm install dash g at gitlawb slash openclaude, then pick a model from twenty-four options. For zero cost, point it at a model running through Ollama on your own computer, no account or key required.

James: The catches are real, though. A local model on consumer hardware is slower and weaker than a hosted frontier model, so it suits smaller, well-scoped tasks. And GitHub flags the repository's license status as unresolved, NOASSERTION, worth a legal look before commercial use.

Maya: Fair point, James, and no public benchmark compares OpenClaude's output against Claude Code's, so the honest claim is cost and choice, not proven quality. That's your Pivot Build daily brief. Check pivotbuild.ai for the full stories. Keep building.