MCP server

Connect an AI client to the Klaracase MCP server (v0.107.0) to ground answers on the corpus and citation graph.

Credits and repeats: what counts as the same answer

Client support

Do you use ChatGPT or Claude in the chat? Go to the guide under AI assistants

The server offers two ways in: a personal bearer token for clients that allow custom headers, and OAuth 2.1 for chat clients that do not. Vendor details as of 2026-11-07.

Claude Code
Works todayAny plan. Add the server with the CLI or a project-scoped .mcp.json.
Cursor
Works todayAny plan. Add the server to ~/.cursor/mcp.json.
VS Code / Windsurf
Works todayAny plan. Both support custom headers on remote MCP servers.
ChatGPT
Works today, via OAuthThe app from the ChatGPT directory works on every plan, including the free one. On Business and Enterprise an administrator may have to approve klaracase for the workspace first. Only paid plans can add a custom app.Guide for ChatGPT
Claude
Works today, via OAuthThe connector from the Claude directory works on every plan, including the free one. On Team and Enterprise an owner may have to approve klaracase for the organization first. According to Anthropic, every plan can also add a custom connector; Free accounts can add one.Guide for Claude
Claude Desktop
Via OAuth, in trialClaude Desktop supports connectors. According to Anthropic, you set them up through your Claude account, not in a configuration file. You do not need a token for this.Guide for Claude

OAuth 2.1 has been live since 2026-07-29. We tested Claude and ChatGPT through to a first tool call with real accounts, ChatGPT through the directory entry. Plan and authentication details from the vendors were checked on 2026-11-07 and can change at any time.

Get a token

Create an API token
For clients that allow a custom Authorization header. Create and revoke your tokens in account settings. A token is shown once. Keep it like a password. If you connect via OAuth you need no token.
Open API tokens in settings

You need to be signed in to open this page. Calls are charged to your organisation's credit balance.

Connect a client

The MCP endpoint is https://klaracase.de/mcp (streamable HTTP, POST). It accepts both a bearer token and OAuth 2.1.

Connect a chat client without a token (OAuth)
For ChatGPT and Claude. No terminal and no token required: use the server address https://klaracase.de/mcp.
  1. Add a custom connector in the client and paste the server address. Leave client ID and secret empty.
  2. Sign in to klaracase, then approve access on the next page.
  3. After you approve, the client takes you back and lists the klaracase tools.

Calls made this way are charged to your organisation's credit balance, exactly like calls made with a token.

Note: The first attempt sometimes fails: after you approve, the browser shows an error page from the chat provider and the connector stays disconnected. That happens on the provider's redirect back, not at klaracase. Your sign-in and approval already reached us. Add the connector again. If that does not help, try a browser window without extensions (a private window). Password managers and similar extensions occasionally interfere with the redirect.

Claude Code (CLI)
Add the server with a custom Authorization header.
claude mcp add --transport http klaracase https://klaracase.de/mcp \
  --header "Authorization: Bearer YOUR_TOKEN_HERE"
Claude Code (.mcp.json)
Or commit a project-scoped .mcp.json to your repo root.
{
  "mcpServers": {
    "klaracase": {
      "type": "http",
      "url": "https://klaracase.de/mcp",
      "headers": {
        "Authorization": "Bearer YOUR_TOKEN_HERE"
      }
    }
  }
}
Cursor
Add to ~/.cursor/mcp.json.
{
  "mcpServers": {
    "klaracase": {
      "url": "https://klaracase.de/mcp",
      "headers": {
        "Authorization": "Bearer YOUR_TOKEN_HERE"
      }
    }
  }
}
mcp-remote (stdio bridge)
For clients that only speak stdio, bridge to the remote server with mcp-remote. Claude Desktop does not need this bridge: it uses the connector from your Claude account.
{
  "mcpServers": {
    "klaracase": {
      "command": "npx",
      "args": [
        "mcp-remote",
        "https://klaracase.de/mcp",
        "--header",
        "Authorization: Bearer YOUR_TOKEN_HERE"
      ]
    }
  }
}

A connection is working once the client lists the klaracase tools and a first call returns a result. A client that shows the server as connected but lists no tools has not authenticated. See below.

Working with AI assistants

An AI assistant does not always call klaracase on its own. When it answers from memory, call klaracase yourself, as the rows below show. This makes the answers easier to rely on: you get decisions with sources instead of statements without proof.

ClaudeName klaracase in your question. In most chats Claude picks the connector on its own. If an answer comes without a source, write something like: "Use klaracase."
ChatGPTType @klaracase in the message box and pick the app from the list. This applies to the app from the ChatGPT directory and to an app you added yourself.

Why it matters

Every file number (Aktenzeichen) klaracase returns exists in the corpus with its source. A file number that did not come from klaracase has no source there. When the assistant calls klaracase, you get decisions you can open, cite and check.

Three tips

  • Call klaracase yourself, as described in the rows above.
  • Ask for sources: "with citations".
  • If an answer names a decision without a link, ask the assistant to check it with klaracase. For a whole text, ask it to check the citations.
  • For developers: for a whole text the assistant calls the check_citations tool.

When it does not work

After approving, the chat provider shows an error page and the connector is not connectedAdd the connector again. The second attempt almost always works. The failure happens after our part. Sign-in and approval already succeeded, but the redirect back to the chat provider failed before it collected access. If the problem persists, open the chat provider in a private window with no extensions. You were not charged, and no partial connection remains.
The client cannot connect, or shows the server as connected but lists no toolsThe token is missing, mistyped or revoked. Check that the Authorization header reads "Bearer " followed by the whole token, including the part before the "|". Create a fresh token in settings if in doubt.
A call returns an error message instead of resultsFailures return errors, not empty results. An error means something went wrong on our side, not that the search found nothing. Retry once. If the error persists, send us the message text.
The answer says there are not enough creditsThe call did not run, and klaracase charged nothing. The message lists the cost and your balance. Paid plans renew monthly. Free plan credits do not renew. You can view your usage in the app.
The client gives up before an answer arrivesA long search can exceed a chat client's own timeout. Ask again more narrowly, with a court or a date range, so there is less to retrieve.
There is no option to add a custom connectorYour plan or client does not offer custom MCP connectors. Check the table above and use one of the supported clients.

Credits, API budget and limits

Every priced tool call is paid in credits. This section explains where the credits come from and what a client does when a call is refused.

Plan credits

Each plan includes credits for one credit period. They pay for the web app, the API and MCP together. A paid plan renews them when the next credit period starts.

API budget

On the Professional plan, and on Enterprise where the contract provides it, an organization admin can load a prepaid API budget in euros. The API budget pays for MCP and API calls once the plan credits are used up. It does not expire. The web app does not use it: when the plan credits are used up, the web app stops. The API budget is for AI assistants and programs, which should not stop halfway through a task.

Top up the API budget

Order of payment

The plan credits are spent first. The API budget pays for the rest, at the rate on the pricing page. A call is paid in full or it does not run: if the plan credits and the API budget together cannot pay for it, nothing is executed and nothing is charged. A call never runs in part.

Limits per connection and per token

An organization admin can set a credit limit for each MCP connection (the sign-in of an AI client) and each API token. The limit counts plan credits and API budget credits together in one credit period. When a connection reaches its limit, only that connection is refused. The other connections of the organization keep working. The count starts again with the next credit period.

Three refusals and what a client does

Each refusal is a tool error with a stable error_code and a message. Nothing ran and nothing was charged. A refusal stays the same until a person acts or the credit period ends, so do not retry it in a loop.

error_code: insufficient_credits, reason: api_budget_emptyThe organization may load an API budget, and what it holds does not cover this call. The client stops and tells its user that an organization admin must top up the API budget in the klaracase web app. The refusal carries no link.
error_code: insufficient_credits, reason: api_budget_not_availableThe organization has no API budget that can pay for the call: its plan has none, or its budget is on hold. The client stops and tells its user. The call can run again when the next credit period starts, when an organization admin has changed the plan, or when klaracase has settled the open payment question with the organization.
error_code: token_limit_reachedThe connection or token reached its credit limit. The refusal carries cap, spent and resets_at. The client stops and tells its user that an organization admin must change the limit, or it waits until resets_at. Other connections keep working.

The value plan_credits_exhausted for reason is reserved. The server does not serve it today. A client must tolerate a reason value it does not know and treat it like api_budget_not_available.

A client or an AI agent never tops up, pays or changes a limit by itself. It tells its user what to do and stops.

Fields in the answer

Every answer and every refusal carries the credits block with metered, cost and credits_remaining. cost counts the credits that the plan and the API budget paid together. credits_remaining is the plan credits left and never includes the API budget. An organization that holds an API budget, or may load one, also gets api_budget with cents (what is left, in euro cents), rate_cents_per_credit and credits_at_rate (how many more credits that pays for). For any other organization the key is absent.

Price

The price per credit from the API budget stands on the pricing page. This page shows no price.

Go to the pricing page

Corpus coverage

The corpus holds German federal court decisions and statutes as well as EU instruments. It is not complete, and coverage differs by court and by year. A Fundstelle we cannot resolve is a statement about our coverage, not a judgement about the citation.

See what the Bestand holds

Data processing

The tools process your requests on our server in Germany. When you search by content, external AI services receive the text of a search query and process it in the EU. On cross_type_search the laws section can also send the query to a language model in the EU; bridge: false prevents that. check_citations examines texts without AI services. The privacy and security page names the providers.

Privacy and security

Examples by area of law

MCP tools

Each tool description below lists its purpose, arguments, refusal codes, and credit price. Response fields are documented in the tool guide so a client does not load them on every connection.

Read the MCP tool guide