Claude Code Feature Availability

Which Claude Code features work with a QCode API key and which do not — each row labelled as measured or inferred

Updated 2026-11-03

Need an API key? Sign up →Endpoints & base URLs →

On This Page

Part of Claude Code runs in the local CLI; another part is a hosted service tied to an Anthropic account. With a third-party API key (QCode included) the former works normally and the latter does not. This page draws the line so you do not spend an afternoon debugging.

About the evidence column: rows are marked measured (we actually ran it) or inferred (judged from how the product works, not tested). Do not treat inference as fact — if a row matters to you, verify it yourself.

The table

Feature With a QCode key Evidence
CLI core: chat, file read/write, Bash, MCP, Skills, Plugins, subagents ✅ Measured: claude -p returns normally; file creation succeeds
Permission mode bypassPermissions ✅ Measured: the same task wrote to disk
Permission modes acceptEdits / manual / plan / dontAsk ✅ Inferred: acceptEdits / manual / dontAsk decide locally, no classifier; plan sends shell commands to the classifier while planning when auto mode is available (official)
Permission mode auto ✅ Measured (2026-10-07, Claude Code 2.1.282 / 2.1.292): Write and Bash run normally; dangerous commands such as uploading a private key, deleting a directory outside the working directory, or downloading a remote script and running it are blocked by the classifier. Classifier requests count toward your usage — see "How auto mode works" below. The "all denied" result recorded on the older version of this page (2026-09-18) does not reproduce on current versions
CLAUDE.md (project rules file) ✅ Measured (2026-11-03, Claude Code 2.1.317): rules written in the project's CLAUDE.md take effect in claude -p replies
Hooks ✅ Measured (2026-11-03): a PostToolUse hook configured in the project's .claude/settings.json fires after Write
Image input (vision) ✅ Measured (2026-11-03): pasting, dragging, an @ path reference, and the model reading an image file itself with Read — claude-opus-5-5, claude-sonnet-5-5, claude-haiku-5-5 and claude-sonnet-5 all read the text in the image; see Image Input
/loop (re-run on an interval in the session) ✅ Measured (2026-11-03): /loop 1m … is turned into a cron scheduled task and runs every minute; it lives only in the current session
Prompt caching ✅ Measured (2026-11-03): calling the Messages API directly, the same request with cache_control on system sent twice returns usage.cache_read_input_tokens above 0 the second time
claude ultrareview / /ultrareview ❌ Measured: returns Ultrareview could not launch: Ultrareview is currently unavailable.
claude setup-token ❌ Measured: the CLI's own help says requires Claude subscription
Cloud sessions / Claude Code on the web / mobile ❌ Official: Slack and cloud sessions always go straight to the Anthropic API, never through a gateway — see llm-gateway-connect
Remote Control ❌ Official: unavailable with a gateway configuration active, and disabled whenever ANTHROPIC_BASE_URL points at a non-Anthropic host — same doc
Cloud scheduled tasks /schedule ❌ Official: requires signing in with a claude.ai account; not reachable with a Console API key or from a third-party provider — see feature-availability
Voice dictation ❌ Official: relies on claude.ai identity and its transcription endpoint; unavailable while ANTHROPIC_API_KEY / ANTHROPIC_AUTH_TOKEN is active — same doc
Artifacts (publishing session output to claude.ai) ❌ Inferred: a claude.ai account feature
Self-hosted environments ❌ Inferred: a Team / Enterprise plan feature
Claude in Chrome extension ❌ Inferred: tied to an Anthropic plan
Claude Code Desktop (local sessions) ⚠️ untested Inferred: the desktop app drives the same local CLI and should inherit the same config

How auto mode works

In auto mode a classifier decides whether each action may run: clearly out-of-bounds actions (leaking credentials, deleting data outside the working directory, downloading and running external scripts, and so on) are blocked, everything else runs without a prompt.

  • auto is the default: since Claude Code 2.1.283, interactive sessions start in auto. Since 2.1.285, claude -p and the Agent SDK also start in auto in sessions that can't fetch feature flags. When you connect to QCode with ANTHROPIC_AUTH_TOKEN, Claude Code has no credential to fetch them (the same holds for CI containers that start fresh every run), so claude -p starts in auto whenever no permission mode is set. If your model doesn't support auto (e.g. Haiku), the session falls back to manual mode.

  • Reviews run as Claude Code's own classifier requests: Claude Code first asks the server to review actions as part of its model requests; QCode doesn't return server-side review results, so Claude Code sends its own classifier request for each action that needs review (default model claude-sonnet-5). These requests count toward your usage. Reads, searches and edits inside the working directory (outside protected paths such as .git and .claude) skip the classifier; subagents are also reviewed when they start and when they report back (official behavior). In your dashboard usage they appear as claude-sonnet-5; how to check and reduce them: Why Claude Sonnet 5 Appears in Your Usage.

  • You may see a notice: when Claude Code switches to its own classifier requests it shows We're changing auto mode to no longer charge for classifier requests … your requests go through <your endpoint host>, which isn't compatible with this update. In an interactive session it holds the first action that needs review until you press Enter (Esc cancels that action); once acknowledged it doesn't reappear on that machine for 24 hours. With claude -p it is only printed to stderr and nothing waits. auto keeps working. Set CLAUDE_CODE_AUTO_MODE_SERVER=0 and the notice no longer appears (see the setup below). What the notice means and how to handle it: Claude Code Classifier Billing Notice.

Don't want auto? Press Shift+Tab in a session, pass --permission-mode at launch, or set a default in ~/.claude/settings.json:

# day to day: edits apply, commands still prompt
claude --permission-mode acceptEdits

# CI: only actions that need no approval and pre-approved tools run; everything else is denied (set the mode explicitly)
claude -p "..." --permission-mode dontAsk --allowedTools "Read,Edit,Bash(npm test)"

# isolated environment (use with care)
claude --permission-mode bypassPermissions
{ "permissions": { "defaultMode": "acceptEdits" } }

Pointing Claude Code at QCode

export ANTHROPIC_BASE_URL="https://api.qcode.cc/api"
export ANTHROPIC_AUTH_TOKEN="cr_your_qcode_key"
export CLAUDE_CODE_AUTO_MODE_SERVER=0   # auto mode uses Claude Code's own classifier requests directly; removes the notice above
claude -p "reply with exactly: OK"

OK means you are connected. See Environment Variables.

This table will go stale

Claude Code ships weekly and the boundary moves. If you observe behaviour that contradicts this page, trust your own measurement, and please tell us at hi@qcode.cc so we can update it.

Get started with QCode

One API key for Claude Code, Codex and more — billed per token, per-model prices on qcode.cc/models

Enterprise →