
What happened
A developer released cameo, a roughly 400-line Go program that reads a TOML file, starts a local HTTP proxy, and relaunches Claude Code with third-party models registered as subagents.
Why it matters
Request routing keys only on the model field, so Claude-bound traffic is forwarded untouched to Anthropic while cheap agents like DeepSeek and GLM go to their own endpoints, which appears to cut token costs.
What to watch
The author dropped Codex support because Codex has fully moved to the Responses API while most of these models still only offer the older Chat Completions API.
WHO IT HITSIndividual developers and small teams on metered Claude Code subscriptions who already hold spare API credits or cheap third-party model quotas can offload routine exploration and multi-step tasks to cheaper subagents.
Summaries like this, in your inbox every morning.
The author's starting problem was uneven supply: Claude Code's subscription tokens run out quickly even though Claude Fable 5.1 and Opus 5.5 are, in the author's words, good enough to call SOTA, while spare capacity sits unused on other models. Free MiMo quota from Xiaomi, occasional cheap access to other Chinese models, and a friend's abandoned GLM subscription were all available but awkward to reach.
Two earlier attempts failed for different reasons. The first, driving another harness through the TUI tool herdr, worked but let the two harnesses approve each other's permission checks, bypassing the permission system entirely and sitting outside Claude Code's permission mechanism and Codex's sandbox. The second, wrapping the author's own harness paimon behind a skill, never convinced Claude Code to treat paimon as a true subagent, and tracing paimon's session history cost so many steps that simple tasks may have burned more tokens than they saved.
The breakthrough was recognizing that many of these cheap models are Anthropic API compatible, so no protocol translation is needed — the same reason the author skipped Claude Code Router as more feature-rich and complex than required. Codex support was deliberately dropped, since Codex has fully moved to the Responses API while most of these models only offer the older Chat Completions API.
Pick your industry and the AI tools you use, and get news related to your work every day.
Free · 30 seconds with Google · unsubscribe anytimeWhat is AIToday? →
Ask AI anything about this article. The AI reads this article, earlier AIToday articles, and Wikipedia, and cites its sources. Q&As are published on this page for other readers too.
Manulife Financial Corporation launched a first-of-its-kind CoverMe travel insurance plugin in ChatGPT in late…

Google said free Gemini app users will be restricted to the "Flash-Lite" model starting October 9; "Flash" nee…

Indie developer Robert Varadan argues that AI models like Opus 5.5 and GPT-6 Astra can clone game demos from a…

Engineer Takashiba, who works on business systems, built a shogi AI on a single RTX 3060 using the dlshogi app…

A developer published Rai, a small Rust engine that runs language models on an ordinary PC's CPU, using only t…

The developer published fk2000/minecraft-ai-bot, which uses Jev to assemble instructions, Gemini to pick from…
