Skip to main content

Listed

, home
FREN

Free API and MCP

Everything Vuntum publishes about robots (profiles, sourced claims, prices, availability, comparisons, changes) can be read by a program or by an AI assistant, with no key, no sign-up and nothing to pay. Every value comes with its source, its level of evidence and its date.

The essentials

  • Free: no key and no account for everyday use. A pro access, with a key, exists for commercial use and heavy volumes.
  • CC BY-NC 4.0 licence: free non-commercial reuse of the factual data, with attribution required (see the terms).
  • REST API: https://vuntum.com/api/v1, read-only, JSON, 120 requests per minute and 300 requests per day per IP address. OpenAPI 3.1 description: openapi.json.
  • MCP server: https://vuntum.com/mcp (Streamable HTTP), read-only, 20 tools, 60 requests per minute and 100 tool calls per day per IP address. Nothing to install.
  • Two languages: texts are in French by default, in English with ?lang=en (API) or locale: "en" (MCP, accepted by every tool, which then also return English page addresses). The descriptions of the MCP tools are written in English, as the protocol expects.

Sample API requests

Each example can be pasted as is in a terminal. Every response has the same shape: data (the result), meta (date, licence, attribution) and, for lists, a pagination cursor. Add limit= to shorten a list.

Find robots

  • Toy robots

    curl 'https://vuntum.com/api/v1/robots?segment=toy&limit=5'
  • Humanoid robots

    curl 'https://vuntum.com/api/v1/robots?category=humanoide&limit=5'
  • Housework robots

    curl 'https://vuntum.com/api/v1/robots?utility=menage&limit=5'
  • Robots that climb stairs (validated claim)

    curl 'https://vuntum.com/api/v1/robots?capability=escaliers&limit=5'
  • Programmable robots (public SDK, Python, ROS)

    curl 'https://vuntum.com/api/v1/robots?programmable=true&limit=5'
  • Search by word

    curl 'https://vuntum.com/api/v1/search?q=unitree&type=robot&limit=5'

Read a profile and its evidence

  • A robot’s full profile

    curl 'https://vuntum.com/api/v1/robots/unitree-g1'
  • All its claims, with source, quote and level of evidence

    curl 'https://vuntum.com/api/v1/robots/unitree-g1/claims'
  • A manufacturer’s profile

    curl 'https://vuntum.com/api/v1/manufacturers/unitree'

Compare

  • Two robots side by side

    curl 'https://vuntum.com/api/v1/compare?robots=unitree-g1,unitree-h1'
  • Up to four robots

    curl 'https://vuntum.com/api/v1/compare?robots=unitree-g1,unitree-h1,unitree-r1-air'

Prices and availability

  • Prices seen on the US market

    curl 'https://vuntum.com/api/v1/prices?market=US&limit=10'
  • Availability on the US market

    curl 'https://vuntum.com/api/v1/availability?market=US&limit=10'
  • Official list prices, with their source

    curl 'https://vuntum.com/api/v1/claims?predicate=price.msrp&limit=10'

Follow what changes

  • The log of publications, corrections and price changes

    curl 'https://vuntum.com/api/v1/changes?limit=20'
  • The observatory’s statistics

    curl 'https://vuntum.com/api/v1/observatory'

Every endpoint, its parameters and its error codes are described in openapi.json. Responses carry an ETag and accept conditional requests; errors follow RFC 9457.

The MCP server

MCP (Model Context Protocol) is the standard plug that lets an AI assistant query a data source. Once Vuntum is added as a connector, the assistant answers from our sourced data rather than from memory, and points to the profile to read. The address to give is always the same: https://vuntum.com/mcp, with no authentication.

Questions to ask your assistant

A few examples, by theme, that call on the server’s tools.

  • Find
    • Which robots can climb stairs, and who says so?
    • Which home robots can I program in Python or ROS 2?
    • Which open-source humanoid robotics projects does Vuntum list?
  • Compare
    • Compare the Unitree G1 and the Unitree H1: what differs, and what is not disclosed?
    • Among these three robots, which has the best evidence for its battery life?
  • Prices and availability
    • What does the Unitree G1 cost in the US and in China, and when was that price seen?
    • Is this robot available to buy in Europe today?
  • Check
    • Where does this battery figure come from, and is it claimed by the maker or measured by a third party?
    • Explain how Vuntum ranks its levels of evidence.
  • Follow
    • What changed in the database this week?
    • Which recent Vuntum articles are about humanoid robots?

MCP clients: what we tested

We separate what we ran ourselves from what we copy from the client’s documentation. The tests were run from a server on 3 and 28 October 2026, against the server’s public address (not a local copy).

Tested by us

Claude Code

Tested: works. The server was added to Claude Code, which called the price tool on the Unitree G1 and answered with both prices (China and US) and their date. One command is enough:

claude mcp add --transport http vuntum https://vuntum.com/mcp

Or, for a project, in the .mcp.json file:

{
  "mcpServers": {
    "vuntum": {
      "type": "http",
      "url": "https://vuntum.com/mcp"
    }
  }
}

Codex CLI (OpenAI)

Setup tested: accepted. Codex registers the server as a Streamable HTTP server and shows it as enabled, with no authentication to provide. We also replayed the protocol handshake and a tool call by hand against the public server, with no error; we did not have a Codex model answer. One command is enough:

codex mcp add vuntum --url https://vuntum.com/mcp

Or, by hand, in ~/.codex/config.toml:

[mcp_servers.vuntum]
url = "https://vuntum.com/mcp"

Gemini CLI (Google)

Tested: works. Gemini CLI connects to the server and reports it as “Connected” (gemini mcp list). It disables servers in a folder it does not trust: run it in a trusted folder. One command is enough:

gemini mcp add --scope user --transport http vuntum https://vuntum.com/mcp

Or, by hand, in ~/.gemini/settings.json:

{
  "mcpServers": {
    "vuntum": {
      "url": "https://vuntum.com/mcp",
      "type": "http"
    }
  }
}

Claude Desktop, through the configuration file (mcp-remote bridge)

Bridge tested: works. Claude Desktop reads this file to start a local program; the mcp-remote bridge connects that program to the remote server. We tried the bridge itself (16 tools listed, one successful call), but not the Claude Desktop application. Node.js is required. In Settings > Developer > Edit Config, add:

{
  "mcpServers": {
    "vuntum": {
      "command": "npx",
      "args": [
        "-y",
        "mcp-remote",
        "https://vuntum.com/mcp"
      ]
    }
  }
}

Your own program (official SDK)

Tested: works. With the official TypeScript SDK (@modelcontextprotocol/sdk), twenty lines are enough to list the tools and call one. A ready-to-run example and a check are in the server’s public repository.

Documented setup, not yet tested by us

We cannot open these applications from our server. The settings below follow their documentation; if one does not work for you, tell us and we will correct the page.

claude.ai and Claude Desktop (custom connector)

In Customize > Connectors, choose Add custom connector, paste the address https://vuntum.com/mcp, leave authentication empty, then confirm. In a conversation, turn it on from the + menu, under Connectors. On Team and Enterprise plans, an organization owner adds it first.

ChatGPT

Remote MCP servers need developer mode (Settings > Security and login > Developer mode, on plans that offer it). Then create an app with the address https://vuntum.com/mcp and No authentication. Menu names change often at OpenAI: follow its developer mode guide.

Cursor

In ~/.cursor/mcp.json (or .cursor/mcp.json in a project):

{
  "mcpServers": {
    "vuntum": {
      "url": "https://vuntum.com/mcp"
    }
  }
}

VS Code

In .vscode/mcp.json:

{
  "servers": {
    "vuntum": {
      "type": "http",
      "url": "https://vuntum.com/mcp"
    }
  }
}

Limits to know

  • Beyond 60 requests per minute per IP address, the MCP server answers 429 with Retry-After: 60; the API is limited to 120 per minute and 300 per day per address (2,000 for AI assistant connectors, search engines and badge services), with explicit RateLimit-* headers; beyond the daily quota it answers 429 with Retry-After until midnight UTC and the link to the pro access. A client session also counts opening and listing the tools in that per-minute limit.
  • The free access to the MCP server also counts 100 tool calls per day per IP address (UTC day). Opening a session and listing the tools do not count. Beyond that, the server answers 429 with a message giving the link to the pro access, until midnight UTC.
  • Only published data is served. A missing value means we have not found a reliable source yet, not that the answer is “no”.
  • For heavy use, prefer the full export (JSON Lines or CSV) to thousands of requests.

Licence and attribution

The factual data is published under the CC BY-NC 4.0 licence: you may copy, transform and redistribute it for non-commercial use, provided you credit the source. Commercial use needs the pro access. Articles, guides and analyses remain protected.

Expected credit: Source: Vuntum (https://vuntum.com), CC BY-NC 4.0 licence, with a link to the profile concerned.

Terms of use

  • Attribution required. Every answer of the MCP server carries an attribution field, for example “Source: Vuntum, https://vuntum.com/en/robots/g1, CC BY-NC 4.0”, and every API answer a meta.attribution field. It goes with the data wherever the data is shown, with its link.
  • No commercial use without the pro access. The license field of every MCP answer says so (commercialUse), as does the meta.commercialUse field of the API. An AI assistant that quotes our data to someone, with the attribution, is welcome and stays within the free access.
  • Earlier releases. The dataset releases published before 24 October 2026 remain under the CC BY 4.0 licence: these terms apply to the data served since that date.
  • Monthly dataset. To work offline or to cite an exact version, a monthly snapshot of the main characteristics is published on Hugging Face, under the same licence, with a change log and a ready-made citation.

Pro access

For commercial or heavy use, the MCP server and the REST API accept a key: a daily quota set for each key (20,000 tool calls by default), 600 requests per minute and a licence for commercial use of the data, stated in the license field of every answer. The key goes in the Authorization: Bearer <key> or X-Vuntum-Key header. An unknown or revoked key is refused (403), never silently turned into the free access. The real-time full export of the API (/export/full.jsonl and /export/robots.csv) is for pro keys only; for free access, the weekly open dataset on Hugging Face replaces it.

To get a key, fill in the form on the business page or write to contact@vuntum.com, describing the intended use. A key is shown only once, when it is created: we keep a fingerprint of it only and cannot recover it. A lost key is revoked, then replaced.

Webhook alerts (pro key)

A pro key can receive an alert each time the price or availability of a list of robots changes. There is no endpoint to subscribe: the subscription (list of robots, https receiving address, a signing secret of its own) is created by Vuntum at your request, and the secret is handed over once. Without an active subscription on an active pro key nothing is sent; a revoked key stops the deliveries.

Each delivery is a JSON POST carrying one change as get_change_history describes it: id, entityType (robot), entityId, kind (price_changed or availability_changed), robot (slug and name), summary (sentence in French), summaryEn (sentence in English) and createdAt (ISO 8601 date). The headers Vuntum-Event (the kind of change) and Vuntum-Delivery (delivery identifier) come with it. Answer with a 2xx code to acknowledge.

Checking the signature: the Vuntum-Signature header is t=<seconds>,v1=<hex>. Recompute HMAC-SHA256(secret, "<seconds>.<raw body received>") as hexadecimal over the exact body received (before any JSON decoding), compare it with v1 in constant time (crypto.timingSafeEqual in Node, hmac.compare_digest in Python) and reject the delivery if t is more than 5 minutes away from your clock.

5-second timeout per delivery, no redirect followed, https on port 443 only, private or local addresses refused. On failure, 5 attempts in all (1, 5, 30 then 120 minutes apart), then the delivery is abandoned. The same change can arrive twice if a reply is lost: use id to deduplicate.

For the rest (exports, llms.txt, dataset): Data and API. The MCP server is also listed in the protocol’s official registry as com.vuntum/mcp, and its public repository holds the code examples.