Finest for Claude Code

Claude Code through the Finest gateway. Your Claude plan serves until its cap. When the cap hits, the overflow valve continues the same conversation on metered billing through Finest, and flips back when your plan window resets. Every metered request gets a receipt: the model that served, what it cost, and what it saved against the model you asked for.

One-command setup

npx @finest-ai/claude-code

The wizard asks for your key (from https://finest.so/console), which mode you want, and whether to alias claude to the finest launcher. finest off restores your previous setup in one step.

The two lanes

• Finest Frontier (finest-frontier): near frontier quality within a close, published bound, much cheaper. Your requested frontier model serves verbatim unless a sealed evidence cell authorizes a substitution for that task; where one does, the receipt names the served model, the quality tier, and the published bound. Hard turns, plan mode, and compaction always serve the pinned frontier model.
• Finest Workhorse (finest-workhorse): your daily driver, task-adequate against sealed absolute quality floors, much cheaper. The lane every earlier install knows as finest-quick (and finest-fast before that); the old ids serve forever.

Everything else in the catalog is a pin: explicit model ids (claude-opus-5, kimi-k3, deepseek-v4-flash, grok-4.6, ...) serve verbatim to that vendor, never second-guessed. Per-request, x-finest-no-demote: true pins any call.

The evidence behind every substitution is published: sealed, reproducible verification records with pass AND fail verdicts, at https://finest.so/verification.

What it costs

No model markup. You pay the host's published rate for the model that actually served. Finest's fee is 25% of the saving it proves on a request, computed against what your requested model would have cost. No saving, no fee. A request served exactly as you asked carries no fee.

Controls and visibility

• Per-key rpm, token, and spend caps, set in the console.
• The statusline shows the session's spend, the as-asked counterfactual, the swap count, and your remaining credit live inside Claude Code, with a link to the latest receipt.
• Sessions launched with finest claude see more: the launcher opens a 127.0.0.1 pass-through in front of the gateway, so the x-finest-* headers already on every response (served model, quality tier, receipt id) also reach the statusline. The render then shows the last turn's served model with its tier chip and a live as-asked vs substituted count, and prints one honest line at start naming the lane that serves. The relay forwards every byte verbatim, records header facts only, and the launch proceeds unchanged if it cannot start ("observer": false in ~/.claude/finest-config.json turns it off).
• finest session prints the session's serving story from that record: turns served as asked vs substituted, the sealed tier bars behind each substitution, the last receipt, and the money from the session summary when it is cached.
• finest status shows the active config; finest valve on|off arms or disarms the overflow valve; finest desktop on|off switches the Claude Desktop app.

Verify a deployed app

finest verify-app <url> is the last step of a build: it triggers the app once and watches your workspace for the receipt. It prints VERIFIED with the receipt id and link, BYPASSED when the app answered and Finest saw nothing, UNREACHABLE when the trigger got no answer, or INCONCLUSIVE with the reason. Exit codes are 0, 1, 2 and 3 in that order, and the last line of output is a JSON object, so a coding agent can run it and paste the result. Pass the request the app expects (--body '{...}'), a command that drives it (--run '...'), or --watch and use the app yourself. A Claude Code session's served receipts carry a client_session_id and never count as the app's. Refusal receipts carry no client_session_id, so a 402 from a Claude Code session inside the window can be cited as the app's refusal; the output says so when it happens, and giving the app its own key and passing --app-key settles it. A receipt whose outcome is client_abort (the app closed its Finest stream) counts as served when output had already reached the app, and as "stopped" before any output: with an answer that is BYPASSED (the app stopped waiting on Finest and answered from somewhere else), without one it is INCONCLUSIVE. Verify the deployment: npx @finest-ai/claude-code@latest verify-app <url> --body '<request body>'. VERIFIED names the receipt; BYPASSED means the app answered and Finest saw nothing.

What this is not

• There is no ongoing free tier. A one-time starting credit covers your first requests, and the wizard states the live figure before your first test request; beyond it, a workspace funds itself before it serves.
• The savings figure is not a constant. Substitutions happen only where a published evidence record authorizes one, and the receipt is the only claim Finest makes about a request.
• Your Claude login stays on your machine. The gateway never sees or uses it; metered serving uses your Finest key.

Leaving

finest off restores your prior Claude Code configuration (env, statusline, hooks) and removes everything Finest installed: the launcher, the shell PATH and alias lines, the local tool files, and the keys the CLI stored. Your pre-Finest settings backup stays as the record of what was restored. If you switched the Mac app (finest desktop on), run finest desktop off too; the app keeps its own key until then. Pins keep serving verbatim until the moment you leave.

Pointers

• Canonical page: https://finest.so/code
• Agent install and entity reference: https://finest.so/llms.txt
• Running Claude Code past plan limits: https://finest.so/guides/run-claude-code-past-plan-limits
• Verification records: https://finest.so/verification
• Console and keys: https://finest.so/console