Claude code

Claude Code on ppq.ai

ppq.ai exposes a native Anthropic /v1/messages endpoint, so Claude Code talks to it directly — no proxy. Everything is driven by env vars scoped to the launch:

Env var	Meaning	Typical value
ANTHROPIC_BASE_URL	point Claude Code at ppq	https://api.ppq.ai
ANTHROPIC_AUTH_TOKEN	ppq API key	from ~/.config/ppq/api-key (none? → ppq:setup)
ANTHROPIC_MODEL	main model	e.g. moonshotai/kimi-k3
ANTHROPIC_SMALL_FAST_MODEL	background tasks (titles, summaries)	cheap, e.g. mistralai/mistral-nemo
ANTHROPIC_DEFAULT_OPUS_MODEL ANTHROPIC_DEFAULT_SONNET_MODEL	what tier-pinned subagents resolve to	e.g. claude-sonnet-5
ANTHROPIC_DEFAULT_HAIKU_MODEL	haiku-pinned tasks	the cheap model
DISABLE_PROMPT_CACHING	1 — most ppq upstreams don't do Anthropic prompt caching	1

Always pin the tier aliases: otherwise model-pinned subagents (code-review, etc.) fall through to real Opus via ppq at ~10.55/52.75 USD per 1M tokens. (Prices in this file are written without dollar signs deliberately: a literal dollar-digit sequence is a positional-argument placeholder in skill bodies and gets mangled at render time.)

Ready-made launcher

${CLAUDE_SKILL_DIR}/scripts/claude-ppq -m moonshotai/kimi-k3 [claude args…]
${CLAUDE_SKILL_DIR}/scripts/claude-ppq --models [filter]   # catalog with prices/ctx/privacy (no key needed)

All model info comes from --models — id, in/out price per 1M, context length, privacy level. Never hand-roll curl https://api.ppq.ai/… plus python3/node parsing: stock Windows has no python3 (a Store stub that errors), hand-rolled commands aren't pre-approved (permission prompt every time), and the API key must never appear inside a command line.

Telling the user to launch it in their own terminal: on POSIX shells give the bash script path as-is; on Windows the user's terminal is PowerShell/cmd, where the bash script just opens in a text editor — give them instead: powershell -ExecutionPolicy Bypass -File "<skill dir>\scripts\claude-ppq.ps1" -m <model> (same flags, including --models).

Run these exactly as shown (one command, full path copied byte-for-byte — keep Windows backslashes as printed — nothing chained) — these shapes are pre-approved via allowed-tools for as long as this turn lasts. If a literal unsubstituted ${CLAUDE_SKILL_DIR} appears above (older harness), build the path from this SKILL.md's own location instead — the "Base directory for this skill" header, or the directory you read this SKILL.md from — still the launcher, never curl. Your own calls go through the Bash tool with the bash launcher on every OS, including Windows — the .ps1 exists to hand to the user for their own terminal; the only pre-approved way to run it yourself is its read-only --models check, never a session launch.

Key hygiene & cross-skill calls: to check whether a key is configured, run exactly test -r ~/.config/ppq/api-key (pre-approved; exit 0 = key present); for masked config + balance, the setup skill's read-only bash <plugin>/skills/setup/scripts/setup.sh status shape is pre-approved too. NEVER open the key file with Read or cat — the key must never enter the conversation. Anything that changes the account (new, from-*) goes through invoking the ppq:setup skill (Skill(ppq:setup) is pre-approved), not its script. Mid-flow questions (model choice, pricing tradeoffs): use the AskUserQuestion tool, not a turn-ending prose question — its answer returns as a tool result and the grant stays armed. No key configured → invoke ppq:setup in this same turn (Skill(ppq:setup) is pre-approved here).

Defaults (all env-overridable): model $PPQ_MODEL → moonshotai/kimi-k3, small $PPQ_SMALL_MODEL → mistralai/mistral-nemo, subagent tier $PPQ_AGENT_MODEL → claude-sonnet-5. Suggest symlinking it onto PATH if the user wants it permanently: ln -s "${CLAUDE_SKILL_DIR}/scripts/claude-ppq" ~/.local/bin/.

Creating a per-model launcher

Match the user's shell and OS — your environment context usually names the shell already; if not, echo $SHELL is pre-approved:

• fish: a function in ~/.config/fish/functions/claude-<name>.fish with local set -lx exports so nothing leaks into the session; if one already exists (e.g. claude-kimi.fish), read it and copy its shape.
• bash/zsh: a function in the rc file, or a claude-ppq symlink on PATH plus PPQ_MODEL=<model>.
• Windows: a PowerShell function in $PROFILE wrapping claude-ppq.ps1 -m <model>.

Whatever the shell: key from $PPQ_API_KEY or ~/.config/ppq/api-key, model overridable via $PPQ_MODEL, and for a new model reconsider the small/agent models' prices.

Alternative: per-project settings.json

For a project that should always run on ppq, put the same vars in .claude/settings.json instead of a launcher:

{ "env": {
    "ANTHROPIC_BASE_URL": "https://api.ppq.ai",
    "ANTHROPIC_MODEL": "moonshotai/kimi-k3",
    "ANTHROPIC_SMALL_FAST_MODEL": "mistralai/mistral-nemo",
    "ANTHROPIC_DEFAULT_OPUS_MODEL": "claude-sonnet-5",
    "ANTHROPIC_DEFAULT_SONNET_MODEL": "claude-sonnet-5",
    "ANTHROPIC_DEFAULT_HAIKU_MODEL": "mistralai/mistral-nemo",
    "DISABLE_PROMPT_CACHING": "1"
} }

Never write the API key into a settings file that might be committed — ANTHROPIC_AUTH_TOKEN belongs in the environment / launcher / key file.

Picking a model

Run --models [filter] — it prints price (in/out per 1M tokens), context length, and privacyLevel (zdr zero-data-retention > e2e TEE > anon) for every model, keyless. Check price and privacy before recommending. --models is the ONLY way you fetch model data — never curl the API, on any turn. On a turn where the grant has expired (any turn after the one that invoked this skill) the launcher command simply asks for one approval — that single familiar prompt always beats a hand-rolled curl chain, which prompts too and then breaks on parsing. Reference points surveyed 2026-07, in/out USD per 1M tokens: kimi-k3 3.17/15.82 (anon), claude-sonnet-5 2.11/10.55, claude-haiku-4.5 1.05/5.28, deepseek-v4-pro 0.46/0.92 (zdr), glm-5.2 0.82/2.57 (zdr, 1M ctx), qwen3-coder-30b 0.07/0.28 (zdr), grok-4.20 1.32/2.64 (2M ctx).

Known quirks

• kimi-k3 latency can spike to minutes — don't assume a hang.
• Balance exhaustion mid-session returns 402 → ppq:topup skill; keep headroom ≈ 2× the planned spend.
• No Anthropic prompt caching on non-Claude upstreams — that's why DISABLE_PROMPT_CACHING=1; expect full-price input tokens each turn.