CarbonMeld MCP server
Carbon credit data in your AI assistant: connect your assistant or MCP client by signing in with your Premium account, or with your key — project search, ratings, CORSIA registry labels and buyer profiles.
What it does
Give your assistant access to CarbonMeld project records and registry facts through the Model Context Protocol (MCP). The MCP handler runs no language model: your own assistant interprets the returned data. It reads CarbonMeld records and does not issue, buy or transfer carbon credits.
The tools
search_projects
Inputs: query, registry, country, methodology, sector, scoring_sector, developer, status, min_score, claim, sort, limit (1–25; default 10), offset
Search by words (all of them must appear in the name, registry code, country, methodology or sector), with optional standard, country, methodology, sector, scoring sector (the sector a project was rated under: the leaderboards' grouping), developer, analysis status, score and claim filters, sorted by rating or name. Returns the total number of matches and one page of them (next_offset when more remain): identifiers, country and its ISO code, methodology, sector, developer, rating with its status, date and method version, registry status, claim tags and project URL.
get_project
Inputs: id (required); lang, sections (optional)
Full available project record by slug or registry code, with the project page's URL and document links (each document's own page when it has one). Its rating is the document-based rating of methodology v3, with its band and evidence coverage; a page that keeps its v2.2 score has that score, labelled; a page with no number says why. The record also holds the negative facts, vintages and claims, and its score section adds the v3 rating's detail (the dimension values and every rule applied, with its source). The fields of the v2.2 analysis (dimension scores, risk indicators, what would improve) stay in the record as data from the v2.2 analysis; a v3-rated page does not show them. A page read under methodology v3 gives no red flags; a page that keeps its v2.2 score gives them, labelled v2.2. The answer adds what the page says beyond the record: its notices as printed, «How this analysis was made», similar projects and the PDF report's address. lang (en, zh, ja) gives the translated texts in that language; sections returns only the parts asked for. Premium details use the shared project quota.
get_projects
Inputs: ids (required; 1–10 slugs or registry codes); lang, sections (optional)
Several project records in one call: the same record as get_project for each id, in the ids' order, with a per-id error where one id fails. Each record found uses the shared project quota, exactly as get_project does. Ids past the invocation's request limit come back retryable and unopened (no quota used): ask again for those ids.
check_offer
Inputs: text (required); vintage, quantity, seller (optional)
Compare pasted offer text with recorded registration, vintages, quantities, claims and developer. This does not confirm ownership or availability of a specific serial number.
corsia_supply
Inputs: lang (optional)
CORSIA registry labels by registry and phase as the site shows them, with the highest-rated First Phase projects and each registry's average rating, from the projects' current ratings. Balances are unretired units of every vintage, not CORSIA-eligible units.
get_buyer
Inputs: name (required; company name or buyer slug); lang, limit (1–50; default 25), offset (optional)
Public retirement records for an organisation as its CarbonMeld page shows them: tonnes, retirements, the quality of what it retired, its projects with the tonnes retired from each and each project's current rating, and its latest retirements. A retirement beneficiary is not necessarily the purchaser; ambiguous names require clarification.
get_methodology
Inputs: code (required); lang (optional)
One methodology as its CarbonMeld page shows it: identity card with sources; how methodology v3 reads projects of this type (the v2.2 rules that stay the same, the changed ones as the v3 rules they became, and a link to the v2.2 rules as they were); version history and transitions; its projects (a ranking of v3 ratings, or a list); the page's URL.
get_ranking
Inputs: dimension (required: sector, country, standard, methodology); value, lang (optional)
The leaderboards as the site shows them: a dimension's boards with their counts and averages, or one board's projects ranked by rating (at least 5 rated), with the page's URL. Sector boards group projects by the sector they were rated under.
get_standard
Inputs: code (required); lang (optional)
One standard as its CarbonMeld page shows it: identity card with sources; how methodology v3 reads projects of this standard (the v2.2 programme rules that stay the same, the changed ones as the v3 rules they became, and a link to the v2.2 rules as they were); the methodologies it issues and its projects use; its projects (a ranking of v3 ratings, or a list); the page's URL.
get_developer
Inputs: name (required; developer name or slug); lang, limit (1–50; default 25), offset (optional)
One project developer as its CarbonMeld page shows it: its numbers of projects, of projects with a methodology v3 rating and of projects with a registry status of rejected, withdrawn, terminated, suspended or issuance denied, its standards and countries, and its projects in the page's order, each with registry code, country, registry status, its rating as the lists show it (a methodology v3 rating with its band, a v2.2 score labelled v2.2, or «Too early to rate» / «Not yet assessable» / «No public grade applies yet» / «Rating not shown»; marked provisional while re-assessed) or analysis state, and project URL, with the page's URL. There is no company-level rating; ambiguous names require clarification.
ccp_tracker
Inputs: status, program, methodology, lang (all optional)
The ICVCM CCP tracker as the site shows it: the Integrity Council's methodology assessments with their status and the number per status, and for each assessment the number of that programme's projects CarbonMeld tracks under the methodology code (every version: not a count of CCP-labelled credits) and its leaderboard, with the table's source and fetch date.
my_portfolio
Inputs: none
Your own portfolio (My portfolio) as its page shows it: the projects you saved, each with its rating as the lists show it (a methodology v3 rating with its band, a v2.2 score labelled v2.2, or «Too early to rate» / «Not yet assessable» / «No public grade applies yet» / «Rating not shown»; marked provisional while re-assessed), analysis state, registry status, vintages, listings and analysis date; for a project read under methodology v3 its negative facts, for a v2.2-scored one its red flags, labelled v2.2; the dimension values only when every rated project is on the same methodology version; the average of the v3 ratings; two separate counts, the negative facts of the projects read under methodology v3 and the red flags of the v2.2-scored projects; the changes left this week; and the change alerts of the last 30 days for these projects. Read only: saving and removing projects is done on the site. It uses none of the daily project quota.
what_changed
Inputs: view (all, notices, labels), lang, limit (1–100; default 50), offset (all optional)
The last 30 days CarbonMeld recorded: official rule-setter notices (ICVCM, Verra, Gold Standard, SBTi and others; publisher, title, link, day) and registry labels added to or removed from projects (with each project page's URL), plus the count of projects per kind of label change over the whole window.
get_method
Inputs: section, rule, dimension (one at a time; all optional); lang, limit (1–50; default 25), offset (optional)
CarbonMeld's rating method as /methodology/ shows it: methodology v3, the document-based rating. Without arguments: the page's sections, the method's version, the four dimensions with their weights and the address of the archived v2.2 rules. A section: its text as the page shows it, with the rule ids in it and one row for each rule card. A rule by its id (A-START, Q-FNRB-ABOVE) or by the key of its card: its name, what it does, each effect and its figure, since when it applies, its source and the source's own words. A v2.2 id (CM-…) gives that rule as v2.2 had it and its place in v3: the same, changed (with the v3 rule it became) or dropped. A dimension (additionality, quantification, permanence, claim_safety, overall): the rules that act on it, in pages (limit, offset; next_offset). Methodology v3 has no criteria.
search_articles
Inputs: query, tag, from, to (YYYY-MM-DD), sort (newest, oldest), lang, limit (1–50; default 10), offset (all optional)
Search CarbonMeld's articles in English, Chinese or Japanese as the site's article index lists them: by words (each in the title, the summary or a tag), by tag (the articles of the site's page for that tag) and by date. Returns the number of matches and one page of them, newest or oldest first (next_offset when more remain): title, date, summary, tags, category and URL. No match is an empty list, not an error.
search_projects claim values: corsia, corsia_first, ccp, ccp_meth, removal, article6. Labels reflect the records CarbonMeld holds.
Connect by signing in
Claude, ChatGPT, Claude Code, Cursor, VS Code and Microsoft 365 Copilot agents connect to CarbonMeld with a sign-in: no key to paste. You approve the connection on a CarbonMeld page, then the assistant reads your Premium data with the same daily limits as the API. Read only: it cannot change anything or make a payment. All the assistants you connect share one daily request limit.
Server URL
https://carbonmeld.com/api/v1/mcp Configuration examples and MCP Registry entry on GitHub
Name in the official MCP Registry: com.carbonmeld/carbonmeld.
Claude (web, desktop and mobile apps)
- Open Customize → Connectors (claude.ai/customize/connectors), choose «+ Add», then «Add custom connector».
- Name: CarbonMeld. Server URL: https://carbonmeld.com/api/v1/mcp. Choose «Continue».
- Authentication: «Sign in now» (or «Sign in when needed»). OAuth client: «Register automatically».
- Choose «Add». A CarbonMeld page opens: sign in with the email link — open it in the same browser — then approve.
- In a conversation, turn CarbonMeld on with «+ → Connectors».
With your API key instead: in step 3 choose «No sign in», then under «Request headers» add Authorization with the value Bearer followed by your key.
Team and Enterprise plans: a Claude organisation Owner adds the connector in Organization settings → Connectors → Add → Custom → Web; each member then connects with their own sign-in. CarbonMeld is also listed in Claude's connectors directory.
ChatGPT
- Go to chatgpt.com/plugins.
- Select the + button, then «Add custom MCP server».
- Enter a name (CarbonMeld) and a description.
- Under «Connection», enter the server URL https://carbonmeld.com/api/v1/mcp.
- Choose OAuth for authentication: ChatGPT registers itself with CarbonMeld, so no client ID or secret is needed. Review the risk warning and select «I understand and want to continue».
- Select «Create as a plugin», check the tools it lists, and install it.
- Sign in on the CarbonMeld page that opens, then approve. In a new conversation, type @ and select CarbonMeld.
Account and workspace policies apply to adding and using custom MCP servers.
Claude Code
claude mcp add --transport http carbonmeld https://carbonmeld.com/api/v1/mcp Add the server, then run /mcp in Claude Code and choose to authenticate, or run claude mcp login carbonmeld in your shell: the CarbonMeld page opens in your browser; sign in, then approve. If the browser shows a connection error after you approve, paste the full address from its address bar into the prompt Claude Code shows. Add --scope user to use the server in all your code projects.
With your API key instead, add the header. To share the setup in a code project without committing the key, use a .mcp.json that reads it from the environment variable CARBONMELD_API_KEY:
claude mcp add --transport http carbonmeld https://carbonmeld.com/api/v1/mcp \
--header "Authorization: Bearer cmk_YOUR_KEY_HERE" {
"mcpServers": {
"carbonmeld": {
"type": "http",
"url": "https://carbonmeld.com/api/v1/mcp",
"headers": {
"Authorization": "Bearer ${CARBONMELD_API_KEY}"
}
}
}
} Cursor
{
"mcpServers": {
"carbonmeld": {
"url": "https://carbonmeld.com/api/v1/mcp"
}
}
} Add the server to ~/.cursor/mcp.json (every code project) or .cursor/mcp.json (one code project). Cursor asks you to sign in to it: the CarbonMeld page opens; sign in, then approve. Cursor's web app and Cursor Agents can also sign in to CarbonMeld.
With your API key instead, add the header and set CARBONMELD_API_KEY in the environment Cursor starts from:
{
"mcpServers": {
"carbonmeld": {
"url": "https://carbonmeld.com/api/v1/mcp",
"headers": {
"Authorization": "Bearer ${env:CARBONMELD_API_KEY}"
}
}
}
} VS Code (GitHub Copilot)
- Run «MCP: Add Server» from the Command Palette.
- Choose «HTTP (HTTP or Server-Sent Events)» and enter https://carbonmeld.com/api/v1/mcp.
- Enter the server ID carbonmeld, then choose «Global».
- When the server starts, VS Code asks you to sign in: your browser opens the CarbonMeld page; sign in, then approve.
The entry this writes, in the portable file ~/.copilot/mcp-config.json:
{
"mcpServers": {
"carbonmeld": {
"type": "http",
"url": "https://carbonmeld.com/api/v1/mcp"
}
}
} With your API key instead, add a header that reads the key from the environment (a portable file reads environment variables, not prompts) and set CARBONMELD_API_KEY in the environment VS Code starts from. Do not paste the key into a chat or commit it to a code project.
{
"mcpServers": {
"carbonmeld": {
"type": "http",
"url": "https://carbonmeld.com/api/v1/mcp",
"headers": {
"Authorization": "Bearer ${CARBONMELD_API_KEY}"
}
}
}
} Microsoft 365 Copilot (declarative agent)
Microsoft 365 Copilot has no box to add a server: CarbonMeld reaches it as a plugin of a declarative agent, built with the Microsoft 365 Agents Toolkit (VS Code extension, version 6.12.0 or later). Your Microsoft 365 account needs custom app upload and Copilot access enabled; ask your administrator otherwise. This flow has no API-key option.
- In VS Code, open Microsoft 365 Agents Toolkit → «Create a New Agent/App» → «Declarative Agent».
- «Add an Action» → «Start with an MCP Server» → enter https://carbonmeld.com/api/v1/mcp.
- Authentication type: «OAuth (with dynamic registration)»: no client ID or secret is needed.
- Choose a location and a name for the agent.
- In «Accounts», sign in to Microsoft 365; in «Lifecycle», choose «Provision».
- Open m365.cloud.microsoft/chat, select the agent under «Agents» and ask about a project. When asked, choose to sign in to the agent: the CarbonMeld page opens; sign in, then approve.
Le Chat (Mistral): not yet supported.
To disconnect an assistant, open your account page: «Connected assistants» lists each one with a Disconnect button. It stops at once; your API key is not affected.
Or connect with your API key
Any MCP client that lets you set a remote server URL and a static Authorization header can use your API key instead of a sign-in. Client interfaces differ; the clients' sections above show where the header goes.
- Generate your key on your account page with active Premium access.
- Set the remote server URL below and add the Authorization header with your own key.
- Connect, let the client list the tools, then ask it to search for a project before opening its full record.
Connection settings
https://carbonmeld.com/api/v1/mcp
Authorization: Bearer cmk_YOUR_KEY_HERE Keep the key out of shared files: the JSON settings above read it from the environment variable CARBONMELD_API_KEY. Never paste it into a chat or commit it to a code project.
Transport: JSON-RPC 2.0 over HTTP POST, with JSON responses. Opening the server URL with GET returns 405; it is not a web page.
Protocol version: 2025-06-18
Example: list the tools
curl 'https://carbonmeld.com/api/v1/mcp' \
-H 'Authorization: Bearer cmk_YOUR_KEY_HERE' \
-H 'Content-Type: application/json' \
--data '{"jsonrpc":"2.0","id":1,"method":"tools/list"}' Use CarbonMeld with Code Mode
Some MCP clients — for example Cloudflare's Code Mode (Agents SDK) — turn a server's tools into a typed TypeScript API and let the model write code that calls them, instead of calling one tool at a time. CarbonMeld declares an output schema for every tool and returns each successful result as structuredContent as well as JSON text, so these clients get exact types for every call.
The TypeScript declarations (.d.ts), generated from the tool schemas and the OpenAPI document: https://carbonmeld.com/api/v1/carbonmeld.d.ts
Example: find cookstove projects in Kenya with a document-based rating of 5 or more, fetch the top 3 in one batch, and list each one's rating, band, evidence coverage, how many rules raised or lowered it, and its negative facts.
// Cookstove projects in Kenya with a document-based rating of 5 or more, best first — search (free), then open the top 3
const found = await carbonmeld.search_projects({ sector: "cookstoves", country: "KE", min_score: 5, sort: "rating_desc", limit: 10 });
const batch = await carbonmeld.get_projects({ ids: found.results.slice(0, 3).map((p) => p.id), sections: ["score"] });
// each project not yet opened today counts 1 of the daily 30 — batch.quota says what is left
for (const item of batch.results) {
if (!item.ok) continue; // item.error says why this id gave no record (retryable: ask again for it)
const r = item.project.carbonmeld_rating, d = item.project.doc_rating;
// the rating, its band and evidence coverage; how many rules raised / lowered it; its negative facts (often none)
console.log(item.project.project_name, r.rating, r.rating_band, r.evidence_coverage, d.moves, d.negative);
} What counts against the quota: each project record get_projects returns uses one project of the daily project quota, exactly as get_project — a batch of 10 can use up to 10. Each project not yet opened today counts 1 of the daily project quota (30 with Premium, 300 with Pro), so search first (search_projects is free) and open only what the user needs; projects already opened today, owned or gifted count 0. search_projects and the page tools use none of the project quota; every tool call counts towards the daily request limit.
Limits and fair use
- 300 requests per day per key by default, shared by REST and MCP.
- 30 different Premium projects per day, shared across the website, PDF reports, REST and MCP. Reopening the same project that day counts once. Portfolio projects and projects unlocked for you are exempt; public-only records do not consume this project quota.
- With Pro: 1,000 requests per day per key and 300 different projects per day; everything else as above.
- 20 offer checks per day per account, shared with the website offer checker.
Daily counters reset at 00:00 UTC. A burst limit also applies: 120 requests per minute per IP. List and statistics requests count towards the request limit, not the Premium project quota.
get_project and get_projects use the same Premium project quota as the REST API (get_projects: one project per record returned). Tool calls count towards the daily request limit; connecting, listing the tools and other protocol messages do not (they have a separate, much higher limit).
Authentication and rate errors use HTTP status codes. A tool can return isError: true inside a successful HTTP response: read the tool result. A temporary data error is not a negative finding; retry later.
Data and privacy
Your client sends tool arguments, including any offer text you paste, to CarbonMeld. Returned records go back to your client, where your assistant does the reasoning. Share only the information needed for the check and review your client's data settings. CarbonMeld records use of Premium features and Premium project access.
Same key as the REST API
Use the REST guide for direct HTTP requests, list parameters and error codes.
API guide