hanzohanzhe/Grandet

What it claims

Grandet ranks model APIs by listed price across more than 1,800 official, inference and relay providers, using data from grandet.ai. It ships as an Apache-2.0 agent Skill (product version 0.3.0, current build 20261018b, internally "j2") and as an anonymous MCP server. The page states that it ranks by listed price only. It never judges whether a site is "good" and never repeats supplier claims such as "full strength" or "stable".

How it works

A lookup returns the cheapest few options for a model and usage type (chat, coding, long output). Each result gives the site, the token group to choose, the price, its unit and the observation date. USD credit, CNY credit and cash prices are listed separately and never compared with each other. Conditions such as a minimum top-up or members-only access appear as one-line hints. Prices that contradict the site's own text are shown only as reference. Groups locked to a different client are hidden.

Connecting works like this:

1. The user signs up, tops up and creates the token on the site.
2. The user stores the key with a terminal command. Typing is hidden, and the key never enters the chat.
3. Grandet shows the config diff, backs up the file and writes only after the user confirms.
4. It checks the key, group and model name without spending credit. "Undo" restores the previous setup.

flowchart TD
  A[User says connect it] --> B{Which host?}
  B -->|Codex, Claude Code, OpenCode, Gemini CLI, etc.| C[scripts/connect.mjs writes config]
  B -->|Other local apps| D[Field card + docs link]
  B -->|Cloud chat| E[Hands over command to run]
  C --> F[Backup, confirm, verify without spending]

scripts/connect.mjs covers Codex, Claude Code, OpenCode, Kilo Code, Continue, Gemini CLI, Hermes, OpenClaw and WorkBuddy. It can also run as a local MCP server.

Evidence

Starting with build 20261017, each build was compared with its predecessor on the same 48 questions, run in the Hermes agent with deepseek-v4.1-flash and glm-5.3.

• 20261018b passed 18/18 targets. Quotes matching Grandet's data rose from 97.4% to 100%, at 1.04× tokens.
• 20261017 cut bad-parameter errors from 18 to 0. Median answer length fell from 841 to 484 characters.
• A 32-question check of j2 against 20261016c on deepseek-v4.1-flash found Skill use rising from 90% to 100%. API errors per question fell from 0.47 to 0.03.
• An A/B test of the 20261016h description raised the trigger rate on weaker models from 47% to 85%.

Builds before 20261017 were not tested this way.

How to use it

Add the remote MCP server (Streamable HTTP, no key):

{ "mcpServers": { "grandet": { "url": "https://grandet.ai/mcp" } } }

claude mcp add --transport http grandet https://grandet.ai/mcp

Its tools are grandet_find_cheapest_api, grandet_find_free_credits, grandet_resolve_model, grandet_quote_evidence, grandet_glossary and grandet_list_apps. Plain HTTP endpoints are also available: https://grandet.ai/api/v1/board and https://grandet.ai/api/v1/free-credits?q=gemini.

To install the Skill from git (requires Node.js 20+):

git clone --depth 1 --branch skill-v0.3.0-build.20261018b https://github.com/hanzohanzhe/Grandet.git
cd Grandet
node scripts/install-price-finder.mjs --app claude-code --receipt-source checkout

The --app flag also accepts codex, cursor, opencode, gemini-cli, cline, roo-code, kilo-code, windsurf, continue, hermes, openclaw, trae-cn and trae-intl. Without Node, copy skills/grandet-api-price-finder into the app's skill folder, for example ~/.claude/skills/.

Limits

• Telemetry. The installer sends one anonymous install receipt by default. --no-telemetry disables it.
• Tags vs. main. Clone a tag, not main, which can carry untested changes. skill-v0.3.1 was withdrawn and should not be used.
• Offline behaviour. 20261018b falls back to a snapshot from 18 October 2026 (469 models, 174 free-credit offers) when grandet.ai is unreachable. The smaller 20261018 build cannot answer offline.
• Older builds. These may request pages or fields the site has since retired. Builds before 20261016d can fail with EPERM on native Windows.
• Quality and speed. Listed prices say nothing about quality. Questions on quality, speed or downgraded models get only what Grandet has on record.
• Licensing. Catalogue code is AGPL-3.0-only. Supplier data carries no blanket open-data licence.