Claude Code on a Third-Party API: What Works and What Doesn't

Feature-by-feature table (checked 2026-10-08)

Each row reads: feature → through QCode → basis in the QCode docs → official docs. ① 1M context → works on supported models → QCode environment page: 1M is supported on claude-opus-5, claude-sonnet-5, claude-fable-5, claude-opus-4-8 and claude-sonnet-4-6 → official: behind a gateway Claude Code gives each model it recognizes the same window as on the Anthropic API, and Sonnet 5 and later and Opus 4.7 and later get 1M directly; Claude Code can't detect a lower limit enforced on the gateway side. ② Prompt caching → works → QCode: caching applies automatically and the Claude family supports it; unit prices on /models → official: the gateway must forward cache_control unchanged, otherwise there is no error but every turn bills as uncached input. ③ WebSearch → unverified → not in the QCode table → official: WebSearch runs queries against Anthropic's web search backend, and behind a gateway availability depends on the provider the gateway forwards to. ④ MCP servers → work → QCode: tested as part of the CLI core row → official: available on every provider; claude.ai connectors need a subscription sign-in, and tool search is off by default. ⑤ Subagents → work → QCode: tested in the same row → official: available on every provider. ⑥ Agent Teams → unverified → not in the QCode table → official: experimental and disabled by default; enable with CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS=1. ⑦ Image input → works → QCode: a Claude Code session has three ways to give the model an image, with claude-opus-5 and claude-sonnet-5 as the vision models → official: vision and PDF input are model capabilities and work on every provider that offers the model. ⑧ PDF → unverified → not listed separately in the QCode docs → official: as above. ⑨ Auto permission mode → works → QCode: tested on 2026-10-07 with 2.1.282 / 2.1.292, Write and Bash run, dangerous commands are blocked by the classifier, and classifier requests count toward your usage; an English notice containing "isn't compatible" may appear at the start of a session and goes away with CLAUDE_CODE_AUTO_MODE_SERVER=0 → official: behind a gateway Claude Code asks the server for review by default, and when the gateway drops the review it falls back to its own classifier requests; CLAUDE_CODE_AUTO_MODE_SERVER=0 makes it use its own classifier requests from the start. ⑩ /usage subscription limits → not applicable → QCode: check usage on the dashboard's usage statistics page, priced from qcode.cc/models and more accurate than the CLI display → official: /usage shows session cost, plan usage limits and activity stats, and /cost is an alias; while a gateway credential is active, the subscription's usage limits don't apply. ⑪ Claude Code Desktop → untested, needs separate setup → marked untested in the QCode table → official: the desktop app reads its gateway from its own third-party inference configuration, not from ANTHROPIC_BASE_URL or settings.json, and runs local sessions only in gateway mode. ⑫ VS Code extension → works → QCode: point the extension's environment variable at https://api.qcode.cc/api → official: the VS Code and JetBrains extensions work on every provider; put the gateway variables in claudeCode.environmentVariables. ⑬ Cowork → not counted as a QCode use → QCode: treat Cowork according to your Claude plan and organization settings and do not count it as part of your QCode usage → the QCode docs add that the official gateway docs never mention Cowork. ⑭ Remote Control, voice dictation, cloud sessions, Ultrareview, the Chrome extension, Artifacts → unavailable → QCode: tested, quoting the official docs, or inferred (the Chrome extension and Artifacts are inferred) → official: they need a claude.ai sign-in and are out of reach from a third-party provider.