Developers · Updated August 2026
Connect an AI assistant to the atlas over MCP
The atlas exposes its whole catalog as a Model Context Protocol server. Point Claude, Cursor, or any MCP client at one URL and your assistant can search Lovable features, read a full record, list recent launches, and count the catalog, with a canonical link on every result.
The endpoint
One streamable HTTP endpoint, no key, no account. It serves the same published catalog this site renders, so a tool result never disagrees with the page it cites.
https://atlas.dahlingdigital.com/mcpConnect your client
Pick whichever matches your setup. Both register the server as lovable-atlas.
claude mcp add --transport http lovable-atlas https://atlas.dahlingdigital.com/mcp{
"mcpServers": {
"lovable-atlas": {
"type": "http",
"url": "https://atlas.dahlingdigital.com/mcp"
}
}
}Restart the client after editing a config file. The four tools show up under the server name once the handshake completes.
The tools
4 tools
search_features
Search the catalogHybrid search across the atlas. Exact matches on title, category, and status rank first, then weighted keyword hits across tagline, capabilities, use cases, and description. Synonyms expand automatically, so MCP also matches Model Context Protocol and connector.
| Parameter | Type | Notes |
|---|---|---|
| query | string | Free text, up to 200 characters. Omit it to browse by filter alone. |
| category | string | Exact category name, for example Agent, AI Models, or Editor. |
| status | enum | One of GA, Beta, Alpha, Removed, Deprecated. |
| limit | integer | 1 to 20. Defaults to 20. |
Returns Up to 20 records with name, category, status, tagline, release date, canonical URL, the field that matched, and a short excerpt of the matched concept.
get_feature
Read one feature in fullFetch the complete record for a single feature by its slug. This is the tool to call after a search when the model needs capabilities and use cases rather than a summary line.
| Parameter | Type | Notes |
|---|---|---|
| id* | string | Lowercase slug, for example agent-mode. Must match /^[a-z0-9]+(?:-[a-z0-9]+)*$/. |
Returns Name, category, status, release date, pricing, tagline, description, capabilities, use cases, source, and the canonical atlas URL.
list_recent_launches
See what shipped latelyNewest features first, ordered by release date. Use it for changelog questions like what Lovable shipped this quarter.
| Parameter | Type | Notes |
|---|---|---|
| limit | integer | 1 to 50. Defaults to 10. |
| status | enum | One of GA, Beta, Alpha. Omit for all three. |
| sinceDays | integer | 1 to 3650. Only features released in the last N days. |
Returns A dated line per feature plus the same records as structured content.
catalog_stats
Count the catalogAggregate counts with no arguments. Useful as a first call so the model knows the shape of the catalog before it starts searching.
No parameters.
Returns Total feature count, a breakdown by status, a breakdown by category, and the site origin.
* required. Everything else is optional.
What to ask once it is wired up
You do not name tools in the prompt. The model picks them from the descriptions.
“Which Lovable features are still in Beta, and what does each one do?”
search_features with status Beta, then get_feature per slug
“What has Lovable shipped in the last 90 days?”
list_recent_launches with sinceDays 90
“Does Lovable have anything for connecting to external tools over MCP?”
search_features with query mcp, expanded to connector and integration
“How big is the catalog and which category has the most entries?”
catalog_stats
Limits and privacy
- Read-only by construction. Every tool is annotated read-only and idempotent, and none accept a write payload.
- Search returns at most 20 records per call. list_recent_launches returns at most 50.
- Calls are logged for traceability: tool name, optional client id, success, duration, and any error. No prompt text and no result payloads are stored.
- Every field the server returns is already public on this site. There is nothing private behind the endpoint.
Questions
- Do I need an API key to use the Lovable Feature Atlas MCP server?
- No. The endpoint is public, unauthenticated, and read-only. There is no key to provision and no account to create. Point your client at https://atlas.dahlingdigital.com/mcp and the four tools appear.
- Can the MCP server write to the atlas?
- No. Every tool is annotated read-only and idempotent, and none of them accept a write payload. The server reads the same published catalog the website renders, so a tool result never disagrees with the page it cites.
- Which clients work with it?
- Any client that speaks the Model Context Protocol over streamable HTTP. That includes Claude Code, Claude Desktop, Cursor, and Windsurf. Clients that only support stdio servers need a local HTTP bridge.
- What data does the server log?
- One row per tool call: the tool name, an optional client id, whether the call succeeded, how long it took, and any error string. No prompt text and no result payloads are stored. Every field in the catalog is already public on this site.
Start from the catalog
The same records the tools return, rendered as pages.
- Lovable Cloud
The Postgres, auth, storage, and edge function layer this atlas runs on.
- Agent Mode
Autonomous, end-to-end execution of multi-step changes across the stack.
- Lovable API
Programmatic access to build and edit Lovable projects from your own code.
- GitHub Sync
Two-way sync to a real repository you own.