Research with AI and the open API

KOMET Archive is public. Search, complete OCR texts and original scans are available without an archive account or API key. Connect your AI client through MCP or use the HTML search and REST API.

Connect once, use in new chats

MCP is an open interface for AI tools. We host the tools; your client makes them available to its model. Save the server URL below in your account or client, select no authentication, and enable the saved connection in each chat when needed. Projects may also store your research instructions.

Name: KOMET Archiv
Server URL: https://komet-archiv.de/mcp
Transport: Streamable HTTP
Authentication: None / No authentication / No sign in

A remote web connection needs no local server download. Available connectors, activation and image handling depend on the client, plan and workspace policies. Instructions checked on 10 November 2026; follow the linked vendor guides for current menus.

ChatGPT

In ChatGPT web, open Plugins and choose “+” → “Add custom MCP server”.
Enter the name and URL above, choose “No authentication”, and complete the connection dialog.
Install your own plugin and select it in the chat using @.

Official OpenAI guide

Claude

Open Customize → Connectors → + Add → Add custom connector.
Enter the name and URL above, and select “No sign in” if asked.
Enable the saved connector in your chat using “+” → Connectors.

Team and Enterprise workspaces may require an administrator to add the connection.

Official Claude guide

Other clients

Gemini web and Google AI search: public web access is independent of MCP. A web answer does not establish that every archive result was searched. Check whether the particular interface supports custom MCP servers. Gemini CLI has documented MCP support.

Gemini CLI

Add to the existing mcpServers configuration in ~/.gemini/settings.json, preserving other entries:

{"mcpServers":{"komet-archiv":{"httpUrl":"https://komet-archiv.de/mcp","timeout":30000}}}

Gemini CLI guide

Codex and Claude Code

In an existing Codex installation:

codex mcp add komet-archiv --url https://komet-archiv.de/mcp

For a user-wide connection in Claude Code:

claude mcp add --transport http komet-archiv --scope user https://komet-archiv.de/mcp

Codex guide · Claude Code guide

Cursor and VS Code / Copilot

In Cursor’s existing mcp.json:

{"mcpServers":{"komet-archiv":{"url":"https://komet-archiv.de/mcp"}}}

In VS Code, run “MCP: Add Server”, choose HTTP and enter the URL. Enable tools in the agent chat.

Cursor guide · VS Code guide

Search, read and inspect scans

search(query): start the full archive search with names, places or keywords, not a full question. Maximum eight words or 200 characters.
search_archive(query, from_year, to_year, page, per_page, sort): year filters, sorting and further result pages. Follow metadata.next_page until null.
fetch(id): full OCR text, issue date, page and source links.
get_page_image(id): the actual scan as image content. Requires an image-capable client and model. Whole newspaper pages, not automatically extracted photographs. Maximum 8 MiB.

Page IDs use issue_id:page. Counts refer to issues, not individual mentions. All matching pages of returned issues are listed, while only the best five pages per issue have snippets. Fetch all relevant page texts. REST, MCP and website calls share IP limits. Respect HTTP 429 and wait for Retry-After.

Use KOMET Archive to search for Albert Ritter from 1990 to 2020. Follow all result pages, read relevant texts and inspect original scans. Cite the issue, date, page and source URL, and explain which results you actually checked.

Quick search or full archive search?

Quick search uses Typesense to shortlist candidates, then PostgreSQL to verify matches. It is useful for a quick start.

Full archive search searches the public PostgreSQL OCR index without that candidate shortlist. Use it for systematic research, year filters and all matching issues. Both MCP search tools and the REST search use this mode.

Coverage depends on available issues and OCR quality. “Full” describes the search of the existing public index, not perfect recognition of every historical mention. Original texts and scans retain their original language.

REST API and web access

Open a search example. Search supports q, from_year, to_year, page, per_page (1–50) and sort (relevance, date_asc, date_desc).

GET /api/v1/search: full search. Follow links.next; matching_pages lists all matching pages, while hits contains up to five page snippets per issue.
GET /api/v1/issues/{id}: issue metadata and page links.
GET /api/v1/issues/{id}/pages/{page}: complete OCR text, image_url and canonical HTML source URL.

The HTML search also exposes matches, full texts, scans and source links without JavaScript. A web-capable assistant may use it without MCP; verify which sources it actually retrieved.

Machine-readable guides and discovery

Complete text guide · MCP connection description · OpenAPI · llms.txt · Sitemap

A harness may discover this connection and offer it to the user, but tools become available only when the client connects the server under its own approval rules. The JSON description is a KOMET discovery aid, not a universal automatic installation standard. Search indexing does not install a connector.

Fair use and request limits

REST, MCP and website requests share a budget per IP address. Signing in, changing tools or opening more sessions does not reset it. Users behind the same office, mobile or AI-provider IP share these limits.

Requests	Short burst	Per minute
All requests together	30 / 5 s	300
Search	10 / 5 s	60
Text and metadata	–	240
MCP scans	4 / 10 s	30
PDF exports	2 / 10 s	10

HTTP 429 means pause: wait for the Retry-After header (seconds). The JSON body also contains retry_after. Queue later calls and retry after that delay; avoid parallel retry loops. Limits recover automatically, without a permanent IP ban. Scan images served as static files have a separate hosting policy.

Reuse: KOMET Archive noncommercial permission applies only to rights KOMET controls. Public-domain works, statutory permissions and third-party rights remain unaffected. Commercial reuse of protected material needs separate permission.